mirrors/hermes-agent
1
0
Fork 0
mirror of https://github.com/NousResearch/hermes-agent.git synced 2026-04-25 00:51:20 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

docs(website): dedicated page per bundled + optional skill (#14929)

Browse source 
Generates a full dedicated Docusaurus page for every one of the 132 skills
(73 bundled + 59 optional) under website/docs/user-guide/skills/{bundled,optional}/<category>/.
Each page carries the skill's description, metadata (version, author, license,
dependencies, platform gating, tags, related skills cross-linked to their own
pages), and the complete SKILL.md body that Hermes loads at runtime.

Previously the two catalog pages just listed skills with a one-line blurb and
no way to see what the skill actually did — users had to go read the source
repo. Now every skill has a browsable, searchable, cross-linked reference in
the docs.

- website/scripts/generate-skill-docs.py — generator that reads skills/ and
  optional-skills/, writes per-skill pages, regenerates both catalog indexes,
  and rewrites the Skills section of sidebars.ts. Handles MDX escaping
  (outside fenced code blocks: curly braces, unsafe HTML-ish tags) and
  rewrites relative references/*.md links to point at the GitHub source.
- website/docs/reference/skills-catalog.md — regenerated; each row links to
  the new dedicated page.
- website/docs/reference/optional-skills-catalog.md — same.
- website/sidebars.ts — Skills section now has Bundled / Optional subtrees
  with one nested category per skill folder.
- .github/workflows/{docs-site-checks,deploy-site}.yml — run the generator
  before docusaurus build so CI stays in sync with the source SKILL.md files.

Build verified locally with `npx docusaurus build`. Only remaining warnings
are pre-existing broken link/anchor issues in unrelated pages.
This commit is contained in:
Teknium 2026-04-23 22:22:11 -07:00 • committed by GitHub
parent eb93f88e1d
commit 0f6eabb890
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
139 changed files with 43523 additions and 306 deletions
Download patch file Download diff file Expand all files Collapse all files

161
website/docs/user-guide/skills/optional/autonomous-ai-agents/autonomous-ai-agents-blackbox.md Normal file
View file

@ -0,0 +1,161 @@
---
title: "Blackbox — Delegate coding tasks to Blackbox AI CLI agent"
sidebar_label: "Blackbox"
description: "Delegate coding tasks to Blackbox AI CLI agent"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Blackbox
Delegate coding tasks to Blackbox AI CLI agent. Multi-model agent with built-in judge that runs tasks through multiple LLMs and picks the best result. Requires the blackbox CLI and a Blackbox AI API key.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/autonomous-ai-agents/blackbox` |
| Path | `optional-skills/autonomous-ai-agents/blackbox` |
| Version | `1.0.0` |
| Author | Hermes Agent (Nous Research) |
| License | MIT |
| Tags | `Coding-Agent`, `Blackbox`, `Multi-Agent`, `Judge`, `Multi-Model` |
| Related skills | [`claude-code`](/docs/user-guide/skills/bundled/autonomous-ai-agents/autonomous-ai-agents-claude-code), [`codex`](/docs/user-guide/skills/bundled/autonomous-ai-agents/autonomous-ai-agents-codex), [`hermes-agent`](/docs/user-guide/skills/bundled/autonomous-ai-agents/autonomous-ai-agents-hermes-agent) |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Blackbox CLI
Delegate coding tasks to [Blackbox AI](https://www.blackbox.ai/) via the Hermes terminal. Blackbox is a multi-model coding agent CLI that dispatches tasks to multiple LLMs (Claude, Codex, Gemini, Blackbox Pro) and uses a judge to select the best implementation.
The CLI is [open-source](https://github.com/blackboxaicode/cli) (GPL-3.0, TypeScript, forked from Gemini CLI) and supports interactive sessions, non-interactive one-shots, checkpointing, MCP, and vision model switching.
## Prerequisites
- Node.js 20+ installed
- Blackbox CLI installed: `npm install -g @blackboxai/cli`
- Or install from source:
```
git clone https://github.com/blackboxaicode/cli.git
cd cli && npm install && npm install -g .
```
- API key from [app.blackbox.ai/dashboard](https://app.blackbox.ai/dashboard)
- Configured: run `blackbox configure` and enter your API key
- Use `pty=true` in terminal calls — Blackbox CLI is an interactive terminal app
## One-Shot Tasks
```
terminal(command="blackbox --prompt 'Add JWT authentication with refresh tokens to the Express API'", workdir="/path/to/project", pty=true)
```
For quick scratch work:
```
terminal(command="cd $(mktemp -d) && git init && blackbox --prompt 'Build a REST API for todos with SQLite'", pty=true)
```
## Background Mode (Long Tasks)
For tasks that take minutes, use background mode so you can monitor progress:
```
# Start in background with PTY
terminal(command="blackbox --prompt 'Refactor the auth module to use OAuth 2.0'", workdir="~/project", background=true, pty=true)
# Returns session_id
# Monitor progress
process(action="poll", session_id="<id>")
process(action="log", session_id="<id>")
# Send input if Blackbox asks a question
process(action="submit", session_id="<id>", data="yes")
# Kill if needed
process(action="kill", session_id="<id>")
```
## Checkpoints & Resume
Blackbox CLI has built-in checkpoint support for pausing and resuming tasks:
```
# After a task completes, Blackbox shows a checkpoint tag
# Resume with a follow-up task:
terminal(command="blackbox --resume-checkpoint 'task-abc123-2026-03-06' --prompt 'Now add rate limiting to the endpoints'", workdir="~/project", pty=true)
```
## Session Commands
During an interactive session, use these commands:
| Command | Effect |
|---------|--------|
| `/compress` | Shrink conversation history to save tokens |
| `/clear` | Wipe history and start fresh |
| `/stats` | View current token usage |
| `Ctrl+C` | Cancel current operation |
## PR Reviews
Clone to a temp directory to avoid modifying the working tree:
```
terminal(command="REVIEW=$(mktemp -d) && git clone https://github.com/user/repo.git $REVIEW && cd $REVIEW && gh pr checkout 42 && blackbox --prompt 'Review this PR against main. Check for bugs, security issues, and code quality.'", pty=true)
```
## Parallel Work
Spawn multiple Blackbox instances for independent tasks:
```
terminal(command="blackbox --prompt 'Fix the login bug'", workdir="/tmp/issue-1", background=true, pty=true)
terminal(command="blackbox --prompt 'Add unit tests for auth'", workdir="/tmp/issue-2", background=true, pty=true)
# Monitor all
process(action="list")
```
## Multi-Model Mode
Blackbox's unique feature is running the same task through multiple models and judging the results. Configure which models to use via `blackbox configure` — select multiple providers to enable the Chairman/judge workflow where the CLI evaluates outputs from different models and picks the best one.
## Key Flags
| Flag | Effect |
|------|--------|
| `--prompt "task"` | Non-interactive one-shot execution |
| `--resume-checkpoint "tag"` | Resume from a saved checkpoint |
| `--yolo` | Auto-approve all actions and model switches |
| `blackbox session` | Start interactive chat session |
| `blackbox configure` | Change settings, providers, models |
| `blackbox info` | Display system information |
## Vision Support
Blackbox automatically detects images in input and can switch to multimodal analysis. VLM modes:
- `"once"` — Switch model for current query only
- `"session"` — Switch for entire session
- `"persist"` — Stay on current model (no switch)
## Token Limits
Control token usage via `.blackboxcli/settings.json`:
```json
{
"sessionTokenLimit": 32000
}
```
## Rules
1. **Always use `pty=true`** — Blackbox CLI is an interactive terminal app and will hang without a PTY
2. **Use `workdir`** — keep the agent focused on the right directory
3. **Background for long tasks** — use `background=true` and monitor with `process` tool
4. **Don't interfere** — monitor with `poll`/`log`, don't kill sessions because they're slow
5. **Report results** — after completion, check what changed and summarize for the user
6. **Credits cost money** — Blackbox uses a credit-based system; multi-model mode consumes credits faster
7. **Check prerequisites** — verify `blackbox` CLI is installed before attempting delegation

445
website/docs/user-guide/skills/optional/autonomous-ai-agents/autonomous-ai-agents-honcho.md Normal file
View file

@ -0,0 +1,445 @@
---
title: "Honcho"
sidebar_label: "Honcho"
description: "Configure and use Honcho memory with Hermes -- cross-session user modeling, multi-profile peer isolation, observation config, dialectic reasoning, session su..."
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Honcho
Configure and use Honcho memory with Hermes -- cross-session user modeling, multi-profile peer isolation, observation config, dialectic reasoning, session summaries, and context budget enforcement. Use when setting up Honcho, troubleshooting memory, managing profiles with Honcho peers, or tuning observation, recall, and dialectic settings.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/autonomous-ai-agents/honcho` |
| Path | `optional-skills/autonomous-ai-agents/honcho` |
| Version | `2.0.0` |
| Author | Hermes Agent |
| License | MIT |
| Tags | `Honcho`, `Memory`, `Profiles`, `Observation`, `Dialectic`, `User-Modeling`, `Session-Summary` |
| Related skills | [`hermes-agent`](/docs/user-guide/skills/bundled/autonomous-ai-agents/autonomous-ai-agents-hermes-agent) |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Honcho Memory for Hermes
Honcho provides AI-native cross-session user modeling. It learns who the user is across conversations and gives every Hermes profile its own peer identity while sharing a unified view of the user.
## When to Use
- Setting up Honcho (cloud or self-hosted)
- Troubleshooting memory not working / peers not syncing
- Creating multi-profile setups where each agent has its own Honcho peer
- Tuning observation, recall, dialectic depth, or write frequency settings
- Understanding what the 5 Honcho tools do and when to use them
- Configuring context budgets and session summary injection
## Setup
### Cloud (app.honcho.dev)
```bash
hermes honcho setup
# select "cloud", paste API key from https://app.honcho.dev
```
### Self-hosted
```bash
hermes honcho setup
# select "local", enter base URL (e.g. http://localhost:8000)
```
See: https://docs.honcho.dev/v3/guides/integrations/hermes#running-honcho-locally-with-hermes
### Verify
```bash
hermes honcho status # shows resolved config, connection test, peer info
```
## Architecture
### Base Context Injection
When Honcho injects context into the system prompt (in `hybrid` or `context` recall modes), it assembles the base context block in this order:
1. **Session summary** -- a short digest of the current session so far (placed first so the model has immediate conversational continuity)
2. **User representation** -- Honcho's accumulated model of the user (preferences, facts, patterns)
3. **AI peer card** -- the identity card for this Hermes profile's AI peer
The session summary is generated automatically by Honcho at the start of each turn (when a prior session exists). It gives the model a warm start without replaying full history.
### Cold / Warm Prompt Selection
Honcho automatically selects between two prompt strategies:
| Condition | Strategy | What happens |
|-----------|----------|--------------|
| No prior session or empty representation | **Cold start** | Lightweight intro prompt; skips summary injection; encourages the model to learn about the user |
| Existing representation and/or session history | **Warm start** | Full base context injection (summary → representation → card); richer system prompt |
You do not need to configure this -- it is automatic based on session state.
### Peers
Honcho models conversations as interactions between **peers**. Hermes creates two peers per session:
- **User peer** (`peerName`): represents the human. Honcho builds a user representation from observed messages.
- **AI peer** (`aiPeer`): represents this Hermes instance. Each profile gets its own AI peer so agents develop independent views.
### Observation
Each peer has two observation toggles that control what Honcho learns from:
| Toggle | What it does |
|--------|-------------|
| `observeMe` | Peer's own messages are observed (builds self-representation) |
| `observeOthers` | Other peers' messages are observed (builds cross-peer understanding) |
Default: all four toggles **on** (full bidirectional observation).
Configure per-peer in `honcho.json`:
```json
{
"observation": {
"user": { "observeMe": true, "observeOthers": true },
"ai": { "observeMe": true, "observeOthers": true }
}
}
```
Or use the shorthand presets:
| Preset | User | AI | Use case |
|--------|------|----|----------|
| `"directional"` (default) | me:on, others:on | me:on, others:on | Multi-agent, full memory |
| `"unified"` | me:on, others:off | me:off, others:on | Single agent, user-only modeling |
Settings changed in the [Honcho dashboard](https://app.honcho.dev) are synced back on session init -- server-side config wins over local defaults.
### Sessions
Honcho sessions scope where messages and observations land. Strategy options:
| Strategy | Behavior |
|----------|----------|
| `per-directory` (default) | One session per working directory |
| `per-repo` | One session per git repository root |
| `per-session` | New Honcho session each Hermes run |
| `global` | Single session across all directories |
Manual override: `hermes honcho map my-project-name`
### Recall Modes
How the agent accesses Honcho memory:
| Mode | Auto-inject context? | Tools available? | Use case |
|------|---------------------|-----------------|----------|
| `hybrid` (default) | Yes | Yes | Agent decides when to use tools vs auto context |
| `context` | Yes | No (hidden) | Minimal token cost, no tool calls |
| `tools` | No | Yes | Agent controls all memory access explicitly |
## Three Orthogonal Knobs
Honcho's dialectic behavior is controlled by three independent dimensions. Each can be tuned without affecting the others:
### Cadence (when)
Controls **how often** dialectic and context calls happen.
| Key | Default | Description |
|-----|---------|-------------|
| `contextCadence` | `1` | Min turns between context API calls |
| `dialecticCadence` | `2` | Min turns between dialectic API calls. Recommended 1–5 |
| `injectionFrequency` | `every-turn` | `every-turn` or `first-turn` for base context injection |
Higher cadence values fire the dialectic LLM less often. `dialecticCadence: 2` means the engine fires every other turn. Setting it to `1` fires every turn.
### Depth (how many)
Controls **how many rounds** of dialectic reasoning Honcho performs per query.
| Key | Default | Range | Description |
|-----|---------|-------|-------------|
| `dialecticDepth` | `1` | 1-3 | Number of dialectic reasoning rounds per query |
| `dialecticDepthLevels` | -- | array | Optional per-depth-round level overrides (see below) |
`dialecticDepth: 2` means Honcho runs two rounds of dialectic synthesis. The first round produces an initial answer; the second refines it.
`dialecticDepthLevels` lets you set the reasoning level for each round independently:
```json
{
"dialecticDepth": 3,
"dialecticDepthLevels": ["low", "medium", "high"]
}
```
If `dialecticDepthLevels` is omitted, rounds use **proportional levels** derived from `dialecticReasoningLevel` (the base):
| Depth | Pass levels |
|-------|-------------|
| 1 | [base] |
| 2 | [minimal, base] |
| 3 | [minimal, base, low] |
This keeps earlier passes cheap while using full depth on the final synthesis.
**Depth at session start.** The session-start prewarm runs the full configured `dialecticDepth` in the background before turn 1. A single-pass prewarm on a cold peer often returns thin output — multi-pass depth runs the audit/reconcile cycle before the user ever speaks. Turn 1 consumes the prewarm result directly; if prewarm hasn't landed in time, turn 1 falls back to a synchronous call with a bounded timeout.
### Level (how hard)
Controls the **intensity** of each dialectic reasoning round.
| Key | Default | Description |
|-----|---------|-------------|
| `dialecticReasoningLevel` | `low` | `minimal`, `low`, `medium`, `high`, `max` |
| `dialecticDynamic` | `true` | When `true`, the model can pass `reasoning_level` to `honcho_reasoning` to override the default per-call. `false` = always use `dialecticReasoningLevel`, model overrides ignored |
Higher levels produce richer synthesis but cost more tokens on Honcho's backend.
## Multi-Profile Setup
Each Hermes profile gets its own Honcho AI peer while sharing the same workspace (user context). This means:
- All profiles see the same user representation
- Each profile builds its own AI identity and observations
- Conclusions written by one profile are visible to others via the shared workspace
### Create a profile with Honcho peer
```bash
hermes profile create coder --clone
# creates host block hermes.coder, AI peer "coder", inherits config from default
```
What `--clone` does for Honcho:
1. Creates a `hermes.coder` host block in `honcho.json`
2. Sets `aiPeer: "coder"` (the profile name)
3. Inherits `workspace`, `peerName`, `writeFrequency`, `recallMode`, etc. from default
4. Eagerly creates the peer in Honcho so it exists before first message
### Backfill existing profiles
```bash
hermes honcho sync # creates host blocks for all profiles that don't have one yet
```
### Per-profile config
Override any setting in the host block:
```json
{
"hosts": {
"hermes.coder": {
"aiPeer": "coder",
"recallMode": "tools",
"dialecticDepth": 2,
"observation": {
"user": { "observeMe": true, "observeOthers": false },
"ai": { "observeMe": true, "observeOthers": true }
}
}
}
}
```
## Tools
The agent has 5 bidirectional Honcho tools (hidden in `context` recall mode):
| Tool | LLM call? | Cost | Use when |
|------|-----------|------|----------|
| `honcho_profile` | No | minimal | Quick factual snapshot at conversation start or for fast name/role/pref lookups |
| `honcho_search` | No | low | Fetch specific past facts to reason over yourself — raw excerpts, no synthesis |
| `honcho_context` | No | low | Full session context snapshot: summary, representation, card, recent messages |
| `honcho_reasoning` | Yes | medium–high | Natural language question synthesized by Honcho's dialectic engine |
| `honcho_conclude` | No | minimal | Write or delete a persistent fact; pass `peer: "ai"` for AI self-knowledge |
### `honcho_profile`
Read or update a peer card — curated key facts (name, role, preferences, communication style). Pass `card: [...]` to update; omit to read. No LLM call.
### `honcho_search`
Semantic search over stored context for a specific peer. Returns raw excerpts ranked by relevance, no synthesis. Default 800 tokens, max 2000. Good when you need specific past facts to reason over yourself rather than a synthesized answer.
### `honcho_context`
Full session context snapshot from Honcho — session summary, peer representation, peer card, and recent messages. No LLM call. Use when you want to see everything Honcho knows about the current session and peer in one shot.
### `honcho_reasoning`
Natural language question answered by Honcho's dialectic reasoning engine (LLM call on Honcho's backend). Higher cost, higher quality. Pass `reasoning_level` to control depth: `minimal` (fast/cheap) → `low` → `medium` → `high` → `max` (thorough). Omit to use the configured default (`low`). Use for synthesized understanding of the user's patterns, goals, or current state.
### `honcho_conclude`
Write or delete a persistent conclusion about a peer. Pass `conclusion: "..."` to create. Pass `delete_id: "..."` to remove a conclusion (for PII removal — Honcho self-heals incorrect conclusions over time, so deletion is only needed for PII). You MUST pass exactly one of the two.
### Bidirectional peer targeting
All 5 tools accept an optional `peer` parameter:
- `peer: "user"` (default) — operates on the user peer
- `peer: "ai"` — operates on this profile's AI peer
- `peer: "<explicit-id>"` — any peer ID in the workspace
Examples:
```
honcho_profile # read user's card
honcho_profile peer="ai" # read AI peer's card
honcho_reasoning query="What does this user care about most?"
honcho_reasoning query="What are my interaction patterns?" peer="ai" reasoning_level="medium"
honcho_conclude conclusion="Prefers terse answers"
honcho_conclude conclusion="I tend to over-explain code" peer="ai"
honcho_conclude delete_id="abc123" # PII removal
```
## Agent Usage Patterns
Guidelines for Hermes when Honcho memory is active.
### On conversation start
```
1. honcho_profile → fast warmup, no LLM cost
2. If context looks thin → honcho_context (full snapshot, still no LLM)
3. If deep synthesis needed → honcho_reasoning (LLM call, use sparingly)
```
Do NOT call `honcho_reasoning` on every turn. Auto-injection already handles ongoing context refresh. Use the reasoning tool only when you genuinely need synthesized insight the base context doesn't provide.
### When the user shares something to remember
```
honcho_conclude conclusion="<specific, actionable fact>"
```
Good conclusions: "Prefers code examples over prose explanations", "Working on a Rust async project through April 2026"
Bad conclusions: "User said something about Rust" (too vague), "User seems technical" (already in representation)
### When the user asks about past context / you need to recall specifics
```
honcho_search query="<topic>" → fast, no LLM, good for specific facts
honcho_context → full snapshot with summary + messages
honcho_reasoning query="<question>" → synthesized answer, use when search isn't enough
```
### When to use `peer: "ai"`
Use AI peer targeting to build and query the agent's own self-knowledge:
- `honcho_conclude conclusion="I tend to be verbose when explaining architecture" peer="ai"` — self-correction
- `honcho_reasoning query="How do I typically handle ambiguous requests?" peer="ai"` — self-audit
- `honcho_profile peer="ai"` — review own identity card
### When NOT to call tools
In `hybrid` and `context` modes, base context (user representation + card + session summary) is auto-injected before every turn. Do not re-fetch what was already injected. Call tools only when:
- You need something the injected context doesn't have
- The user explicitly asks you to recall or check memory
- You're writing a conclusion about something new
### Cadence awareness
`honcho_reasoning` on the tool side shares the same cost as auto-injection dialectic. After an explicit tool call, the auto-injection cadence resets — avoiding double-charging the same turn.
## Config Reference
Config file: `$HERMES_HOME/honcho.json` (profile-local) or `~/.honcho/config.json` (global).
### Key settings
| Key | Default | Description |
|-----|---------|-------------|
| `apiKey` | -- | API key ([get one](https://app.honcho.dev)) |
| `baseUrl` | -- | Base URL for self-hosted Honcho |
| `peerName` | -- | User peer identity |
| `aiPeer` | host key | AI peer identity |
| `workspace` | host key | Shared workspace ID |
| `recallMode` | `hybrid` | `hybrid`, `context`, or `tools` |
| `observation` | all on | Per-peer `observeMe`/`observeOthers` booleans |
| `writeFrequency` | `async` | `async`, `turn`, `session`, or integer N |
| `sessionStrategy` | `per-directory` | `per-directory`, `per-repo`, `per-session`, `global` |
| `messageMaxChars` | `25000` | Max chars per message (chunked if exceeded) |
### Dialectic settings
| Key | Default | Description |
|-----|---------|-------------|
| `dialecticReasoningLevel` | `low` | `minimal`, `low`, `medium`, `high`, `max` |
| `dialecticDynamic` | `true` | Auto-bump reasoning by query complexity. `false` = fixed level |
| `dialecticDepth` | `1` | Number of dialectic rounds per query (1-3) |
| `dialecticDepthLevels` | -- | Optional array of per-round levels, e.g. `["low", "high"]` |
| `dialecticMaxInputChars` | `10000` | Max chars for dialectic query input |
### Context budget and injection
| Key | Default | Description |
|-----|---------|-------------|
| `contextTokens` | uncapped | Max tokens for the combined base context injection (summary + representation + card). Opt-in cap — omit to leave uncapped, set to an integer to bound injection size. |
| `injectionFrequency` | `every-turn` | `every-turn` or `first-turn` |
| `contextCadence` | `1` | Min turns between context API calls |
| `dialecticCadence` | `2` | Min turns between dialectic LLM calls (recommended 1–5) |
The `contextTokens` budget is enforced at injection time. If the session summary + representation + card exceed the budget, Honcho trims the summary first, then the representation, preserving the card. This prevents context blowup in long sessions.
### Memory-context sanitization
Honcho sanitizes the `memory-context` block before injection to prevent prompt injection and malformed content:
- Strips XML/HTML tags from user-authored conclusions
- Normalizes whitespace and control characters
- Truncates individual conclusions that exceed `messageMaxChars`
- Escapes delimiter sequences that could break the system prompt structure
This fix addresses edge cases where raw user conclusions containing markup or special characters could corrupt the injected context block.
## Troubleshooting
### "Honcho not configured"
Run `hermes honcho setup`. Ensure `memory.provider: honcho` is in `~/.hermes/config.yaml`.
### Memory not persisting across sessions
Check `hermes honcho status` -- verify `saveMessages: true` and `writeFrequency` isn't `session` (which only writes on exit).
### Profile not getting its own peer
Use `--clone` when creating: `hermes profile create <name> --clone`. For existing profiles: `hermes honcho sync`.
### Observation changes in dashboard not reflected
Observation config is synced from the server on each session init. Start a new session after changing settings in the Honcho UI.
### Messages truncated
Messages over `messageMaxChars` (default 25k) are automatically chunked with `[continued]` markers. If you're hitting this often, check if tool results or skill content is inflating message size.
### Context injection too large
If you see warnings about context budget exceeded, lower `contextTokens` or reduce `dialecticDepth`. The session summary is trimmed first when the budget is tight.
### Session summary missing
Session summary requires at least one prior turn in the current Honcho session. On cold start (new session, no history), the summary is omitted and Honcho uses the cold-start prompt strategy instead.
## CLI Commands
| Command | Description |
|---------|-------------|
| `hermes honcho setup` | Interactive setup wizard (cloud/local, identity, observation, recall, sessions) |
| `hermes honcho status` | Show resolved config, connection test, peer info for active profile |
| `hermes honcho enable` | Enable Honcho for the active profile (creates host block if needed) |
| `hermes honcho disable` | Disable Honcho for the active profile |
| `hermes honcho peer` | Show or update peer names (`--user <name>`, `--ai <name>`, `--reasoning <level>`) |
| `hermes honcho peers` | Show peer identities across all profiles |
| `hermes honcho mode` | Show or set recall mode (`hybrid`, `context`, `tools`) |
| `hermes honcho tokens` | Show or set token budgets (`--context <N>`, `--dialectic <N>`) |
| `hermes honcho sessions` | List known directory-to-session-name mappings |
| `hermes honcho map <name>` | Map current working directory to a Honcho session name |
| `hermes honcho identity` | Seed AI peer identity or show both peer representations |
| `hermes honcho sync` | Create host blocks for all Hermes profiles that don't have one yet |
| `hermes honcho migrate` | Step-by-step migration guide from OpenClaw native memory to Hermes + Honcho |
| `hermes memory setup` | Generic memory provider picker (selecting "honcho" runs the same wizard) |
| `hermes memory status` | Show active memory provider and config |
| `hermes memory off` | Disable external memory provider |

248
website/docs/user-guide/skills/optional/blockchain/blockchain-base.md Normal file
View file

@ -0,0 +1,248 @@
---
title: "Base"
sidebar_label: "Base"
description: "Query Base (Ethereum L2) blockchain data with USD pricing — wallet balances, token info, transaction details, gas analysis, contract inspection, whale detect..."
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Base
Query Base (Ethereum L2) blockchain data with USD pricing — wallet balances, token info, transaction details, gas analysis, contract inspection, whale detection, and live network stats. Uses Base RPC + CoinGecko. No API key required.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/blockchain/base` |
| Path | `optional-skills/blockchain/base` |
| Version | `0.1.0` |
| Author | youssefea |
| License | MIT |
| Tags | `Base`, `Blockchain`, `Crypto`, `Web3`, `RPC`, `DeFi`, `EVM`, `L2`, `Ethereum` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Base Blockchain Skill
Query Base (Ethereum L2) on-chain data enriched with USD pricing via CoinGecko.
8 commands: wallet portfolio, token info, transactions, gas analysis,
contract inspection, whale detection, network stats, and price lookup.
No API key needed. Uses only Python standard library (urllib, json, argparse).
---
## When to Use
- User asks for a Base wallet balance, token holdings, or portfolio value
- User wants to inspect a specific transaction by hash
- User wants ERC-20 token metadata, price, supply, or market cap
- User wants to understand Base gas costs and L1 data fees
- User wants to inspect a contract (ERC type detection, proxy resolution)
- User wants to find large ETH transfers (whale detection)
- User wants Base network health, gas price, or ETH price
- User asks "what's the price of USDC/AERO/DEGEN/ETH?"
---
## Prerequisites
The helper script uses only Python standard library (urllib, json, argparse).
No external packages required.
Pricing data comes from CoinGecko's free API (no key needed, rate-limited
to ~10-30 requests/minute). For faster lookups, use `--no-prices` flag.
---
## Quick Reference
RPC endpoint (default): https://mainnet.base.org
Override: export BASE_RPC_URL=https://your-private-rpc.com
Helper script path: ~/.hermes/skills/blockchain/base/scripts/base_client.py
```
python3 base_client.py wallet <address> [--limit N] [--all] [--no-prices]
python3 base_client.py tx <hash>
python3 base_client.py token <contract_address>
python3 base_client.py gas
python3 base_client.py contract <address>
python3 base_client.py whales [--min-eth N]
python3 base_client.py stats
python3 base_client.py price <contract_address_or_symbol>
```
---
## Procedure
### 0. Setup Check
```bash
python3 --version
# Optional: set a private RPC for better rate limits
export BASE_RPC_URL="https://mainnet.base.org"
# Confirm connectivity
python3 ~/.hermes/skills/blockchain/base/scripts/base_client.py stats
```
### 1. Wallet Portfolio
Get ETH balance and ERC-20 token holdings with USD values.
Checks ~15 well-known Base tokens (USDC, WETH, AERO, DEGEN, etc.)
via on-chain `balanceOf` calls. Tokens sorted by value, dust filtered.
```bash
python3 ~/.hermes/skills/blockchain/base/scripts/base_client.py \
wallet 0xd8dA6BF26964aF9D7eEd9e03E53415D37aA96045
```
Flags:
- `--limit N` — show top N tokens (default: 20)
- `--all` — show all tokens, no dust filter, no limit
- `--no-prices` — skip CoinGecko price lookups (faster, RPC-only)
Output includes: ETH balance + USD value, token list with prices sorted
by value, dust count, total portfolio value in USD.
Note: Only checks known tokens. Unknown ERC-20s are not discovered.
Use the `token` command with a specific contract address for any token.
### 2. Transaction Details
Inspect a full transaction by its hash. Shows ETH value transferred,
gas used, fee in ETH/USD, status, and decoded ERC-20/ERC-721 transfers.
```bash
python3 ~/.hermes/skills/blockchain/base/scripts/base_client.py \
tx 0xabc123...your_tx_hash_here
```
Output: hash, block, from, to, value (ETH + USD), gas price, gas used,
fee, status, contract creation address (if any), token transfers.
### 3. Token Info
Get ERC-20 token metadata: name, symbol, decimals, total supply, price,
market cap, and contract code size.
```bash
python3 ~/.hermes/skills/blockchain/base/scripts/base_client.py \
token 0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913
```
Output: name, symbol, decimals, total supply, price, market cap.
Reads name/symbol/decimals directly from the contract via eth_call.
### 4. Gas Analysis
Detailed gas analysis with cost estimates for common operations.
Shows current gas price, base fee trends over 10 blocks, block
utilization, and estimated costs for ETH transfers, ERC-20 transfers,
and swaps.
```bash
python3 ~/.hermes/skills/blockchain/base/scripts/base_client.py gas
```
Output: current gas price, base fee, block utilization, 10-block trend,
cost estimates in ETH and USD.
Note: Base is an L2 — actual transaction costs include an L1 data
posting fee that depends on calldata size and L1 gas prices. The
estimates shown are for L2 execution only.
### 5. Contract Inspection
Inspect an address: determine if it's an EOA or contract, detect
ERC-20/ERC-721/ERC-1155 interfaces, resolve EIP-1967 proxy
implementation addresses.
```bash
python3 ~/.hermes/skills/blockchain/base/scripts/base_client.py \
contract 0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913
```
Output: is_contract, code size, ETH balance, detected interfaces
(ERC-20, ERC-721, ERC-1155), ERC-20 metadata, proxy implementation
address.
### 6. Whale Detector
Scan the most recent block for large ETH transfers with USD values.
```bash
python3 ~/.hermes/skills/blockchain/base/scripts/base_client.py \
whales --min-eth 1.0
```
Note: scans the latest block only — point-in-time snapshot, not historical.
Default threshold is 1.0 ETH (lower than Solana's default since ETH
values are higher).
### 7. Network Stats
Live Base network health: latest block, chain ID, gas price, base fee,
block utilization, transaction count, and ETH price.
```bash
python3 ~/.hermes/skills/blockchain/base/scripts/base_client.py stats
```
### 8. Price Lookup
Quick price check for any token by contract address or known symbol.
```bash
python3 ~/.hermes/skills/blockchain/base/scripts/base_client.py price ETH
python3 ~/.hermes/skills/blockchain/base/scripts/base_client.py price USDC
python3 ~/.hermes/skills/blockchain/base/scripts/base_client.py price AERO
python3 ~/.hermes/skills/blockchain/base/scripts/base_client.py price DEGEN
python3 ~/.hermes/skills/blockchain/base/scripts/base_client.py price 0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913
```
Known symbols: ETH, WETH, USDC, cbETH, AERO, DEGEN, TOSHI, BRETT,
WELL, wstETH, rETH, cbBTC.
---
## Pitfalls
- **CoinGecko rate-limits** — free tier allows ~10-30 requests/minute.
Price lookups use 1 request per token. Use `--no-prices` for speed.
- **Public RPC rate-limits** — Base's public RPC limits requests.
For production use, set BASE_RPC_URL to a private endpoint
(Alchemy, QuickNode, Infura).
- **Wallet shows known tokens only** — unlike Solana, EVM chains have no
built-in "get all tokens" RPC. The wallet command checks ~15 popular
Base tokens via `balanceOf`. Unknown ERC-20s won't appear. Use the
`token` command for any specific contract.
- **Token names read from contract** — if a contract doesn't implement
`name()` or `symbol()`, these fields may be empty. Known tokens have
hardcoded labels as fallback.
- **Gas estimates are L2 only** — Base transaction costs include an L1
data posting fee (depends on calldata size and L1 gas prices). The gas
command estimates L2 execution cost only.
- **Whale detector scans latest block only** — not historical. Results
vary by the moment you query. Default threshold is 1.0 ETH.
- **Proxy detection** — only EIP-1967 proxies are detected. Other proxy
patterns (EIP-1167 minimal proxy, custom storage slots) are not checked.
- **Retry on 429** — both RPC and CoinGecko calls retry up to 2 times
with exponential backoff on rate-limit errors.
---
## Verification
```bash
# Should print Base chain ID (8453), latest block, gas price, and ETH price
python3 ~/.hermes/skills/blockchain/base/scripts/base_client.py stats
```

224
website/docs/user-guide/skills/optional/blockchain/blockchain-solana.md Normal file
View file

@ -0,0 +1,224 @@
---
title: "Solana"
sidebar_label: "Solana"
description: "Query Solana blockchain data with USD pricing — wallet balances, token portfolios with values, transaction details, NFTs, whale detection, and live network s..."
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Solana
Query Solana blockchain data with USD pricing — wallet balances, token portfolios with values, transaction details, NFTs, whale detection, and live network stats. Uses Solana RPC + CoinGecko. No API key required.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/blockchain/solana` |
| Path | `optional-skills/blockchain/solana` |
| Version | `0.2.0` |
| Author | Deniz Alagoz (gizdusum), enhanced by Hermes Agent |
| License | MIT |
| Tags | `Solana`, `Blockchain`, `Crypto`, `Web3`, `RPC`, `DeFi`, `NFT` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Solana Blockchain Skill
Query Solana on-chain data enriched with USD pricing via CoinGecko.
8 commands: wallet portfolio, token info, transactions, activity, NFTs,
whale detection, network stats, and price lookup.
No API key needed. Uses only Python standard library (urllib, json, argparse).
---
## When to Use
- User asks for a Solana wallet balance, token holdings, or portfolio value
- User wants to inspect a specific transaction by signature
- User wants SPL token metadata, price, supply, or top holders
- User wants recent transaction history for an address
- User wants NFTs owned by a wallet
- User wants to find large SOL transfers (whale detection)
- User wants Solana network health, TPS, epoch, or SOL price
- User asks "what's the price of BONK/JUP/SOL?"
---
## Prerequisites
The helper script uses only Python standard library (urllib, json, argparse).
No external packages required.
Pricing data comes from CoinGecko's free API (no key needed, rate-limited
to ~10-30 requests/minute). For faster lookups, use `--no-prices` flag.
---
## Quick Reference
RPC endpoint (default): https://api.mainnet-beta.solana.com
Override: export SOLANA_RPC_URL=https://your-private-rpc.com
Helper script path: ~/.hermes/skills/blockchain/solana/scripts/solana_client.py
```
python3 solana_client.py wallet <address> [--limit N] [--all] [--no-prices]
python3 solana_client.py tx <signature>
python3 solana_client.py token <mint_address>
python3 solana_client.py activity <address> [--limit N]
python3 solana_client.py nft <address>
python3 solana_client.py whales [--min-sol N]
python3 solana_client.py stats
python3 solana_client.py price <mint_or_symbol>
```
---
## Procedure
### 0. Setup Check
```bash
python3 --version
# Optional: set a private RPC for better rate limits
export SOLANA_RPC_URL="https://api.mainnet-beta.solana.com"
# Confirm connectivity
python3 ~/.hermes/skills/blockchain/solana/scripts/solana_client.py stats
```
### 1. Wallet Portfolio
Get SOL balance, SPL token holdings with USD values, NFT count, and
portfolio total. Tokens sorted by value, dust filtered, known tokens
labeled by name (BONK, JUP, USDC, etc.).
```bash
python3 ~/.hermes/skills/blockchain/solana/scripts/solana_client.py \
wallet 9WzDXwBbmkg8ZTbNMqUxvQRAyrZzDsGYdLVL9zYtAWWM
```
Flags:
- `--limit N` — show top N tokens (default: 20)
- `--all` — show all tokens, no dust filter, no limit
- `--no-prices` — skip CoinGecko price lookups (faster, RPC-only)
Output includes: SOL balance + USD value, token list with prices sorted
by value, dust count, NFT summary, total portfolio value in USD.
### 2. Transaction Details
Inspect a full transaction by its base58 signature. Shows balance changes
in both SOL and USD.
```bash
python3 ~/.hermes/skills/blockchain/solana/scripts/solana_client.py \
tx 5j7s8K...your_signature_here
```
Output: slot, timestamp, fee, status, balance changes (SOL + USD),
program invocations.
### 3. Token Info
Get SPL token metadata, current price, market cap, supply, decimals,
mint/freeze authorities, and top 5 holders.
```bash
python3 ~/.hermes/skills/blockchain/solana/scripts/solana_client.py \
token DezXAZ8z7PnrnRJjz3wXBoRgixCa6xjnB7YaB1pPB263
```
Output: name, symbol, decimals, supply, price, market cap, top 5
holders with percentages.
### 4. Recent Activity
List recent transactions for an address (default: last 10, max: 25).
```bash
python3 ~/.hermes/skills/blockchain/solana/scripts/solana_client.py \
activity 9WzDXwBbmkg8ZTbNMqUxvQRAyrZzDsGYdLVL9zYtAWWM --limit 25
```
### 5. NFT Portfolio
List NFTs owned by a wallet (heuristic: SPL tokens with amount=1, decimals=0).
```bash
python3 ~/.hermes/skills/blockchain/solana/scripts/solana_client.py \
nft 9WzDXwBbmkg8ZTbNMqUxvQRAyrZzDsGYdLVL9zYtAWWM
```
Note: Compressed NFTs (cNFTs) are not detected by this heuristic.
### 6. Whale Detector
Scan the most recent block for large SOL transfers with USD values.
```bash
python3 ~/.hermes/skills/blockchain/solana/scripts/solana_client.py \
whales --min-sol 500
```
Note: scans the latest block only — point-in-time snapshot, not historical.
### 7. Network Stats
Live Solana network health: current slot, epoch, TPS, supply, validator
version, SOL price, and market cap.
```bash
python3 ~/.hermes/skills/blockchain/solana/scripts/solana_client.py stats
```
### 8. Price Lookup
Quick price check for any token by mint address or known symbol.
```bash
python3 ~/.hermes/skills/blockchain/solana/scripts/solana_client.py price BONK
python3 ~/.hermes/skills/blockchain/solana/scripts/solana_client.py price JUP
python3 ~/.hermes/skills/blockchain/solana/scripts/solana_client.py price SOL
python3 ~/.hermes/skills/blockchain/solana/scripts/solana_client.py price DezXAZ8z7PnrnRJjz3wXBoRgixCa6xjnB7YaB1pPB263
```
Known symbols: SOL, USDC, USDT, BONK, JUP, WETH, JTO, mSOL, stSOL,
PYTH, HNT, RNDR, WEN, W, TNSR, DRIFT, bSOL, JLP, WIF, MEW, BOME, PENGU.
---
## Pitfalls
- **CoinGecko rate-limits** — free tier allows ~10-30 requests/minute.
Price lookups use 1 request per token. Wallets with many tokens may
not get prices for all of them. Use `--no-prices` for speed.
- **Public RPC rate-limits** — Solana mainnet public RPC limits requests.
For production use, set SOLANA_RPC_URL to a private endpoint
(Helius, QuickNode, Triton).
- **NFT detection is heuristic** — amount=1 + decimals=0. Compressed
NFTs (cNFTs) and Token-2022 NFTs won't appear.
- **Whale detector scans latest block only** — not historical. Results
vary by the moment you query.
- **Transaction history** — public RPC keeps ~2 days. Older transactions
may not be available.
- **Token names** — ~25 well-known tokens are labeled by name. Others
show abbreviated mint addresses. Use the `token` command for full info.
- **Retry on 429** — both RPC and CoinGecko calls retry up to 2 times
with exponential backoff on rate-limit errors.
---
## Verification
```bash
# Should print current Solana slot, TPS, and SOL price
python3 ~/.hermes/skills/blockchain/solana/scripts/solana_client.py stats
```

113
website/docs/user-guide/skills/optional/communication/communication-one-three-one-rule.md Normal file
View file

@ -0,0 +1,113 @@
---
title: "One Three One Rule — Structured decision-making framework for technical proposals and trade-off analysis"
sidebar_label: "One Three One Rule"
description: "Structured decision-making framework for technical proposals and trade-off analysis"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# One Three One Rule
Structured decision-making framework for technical proposals and trade-off analysis. When the user faces a choice between multiple approaches (architecture decisions, tool selection, refactoring strategies, migration paths), this skill produces a 1-3-1 format: one clear problem statement, three distinct options with pros/cons, and one concrete recommendation with definition of done and implementation plan. Use when the user asks for a "1-3-1", says "give me options", or needs help choosing between competing approaches.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/communication/one-three-one-rule` |
| Path | `optional-skills/communication/one-three-one-rule` |
| Version | `1.0.0` |
| Author | Willard Moore |
| License | MIT |
| Tags | `communication`, `decision-making`, `proposals`, `trade-offs` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# 1-3-1 Communication Rule
Structured decision-making format for when a task has multiple viable approaches and the user needs a clear recommendation. Produces a concise problem framing, three options with trade-offs, and an actionable plan for the recommended path.
## When to Use
- The user explicitly asks for a "1-3-1" response.
- The user says "give me options" or "what are my choices" for a technical decision.
- A task has multiple viable approaches with meaningful trade-offs (architecture, tooling, migration strategy).
- The user needs a proposal they can forward to a team or stakeholder.
Do NOT use for simple questions with one obvious answer, debugging sessions, or tasks where the user has already decided on an approach.
## Procedure
1. **Problem** (one sentence)
- State the core decision or desired outcome in a single concise sentence.
- Focus on the *what*, not the *how* — no implementation details, tool names, or specific technologies.
- Keep it tight. If you need "and", you're describing two problems.
2. **Options** (exactly three)
- Present three distinct, viable approaches labeled A, B, C.
- Each option gets a brief description, pros, and cons.
- Options should represent genuinely different strategies, not minor variations of the same approach.
3. **Recommendation** (one option)
- State which option you recommend and why, based on the user's context and priorities.
- Be direct — this is your professional judgment, not a hedge.
4. **Definition of Done**
- List the specific success criteria for the recommended option.
- These are concrete, verifiable outcomes — not vague aspirations.
- If the user picks a different option, revise this section to match.
5. **Implementation Plan**
- Concrete steps to execute the recommended option.
- Include specific commands, tools, or actions where applicable.
- If the user picks a different option, revise this section to match.
## Verification
- Response has exactly one Problem sentence.
- Response has exactly three Options (A, B, C) with pros and cons for each.
- Response has a single Recommendation that picks one option with reasoning.
- Definition of Done and Implementation Plan align with the recommended option.
- If the user selects a different option, Recommendation, DoD, and Implementation Plan update accordingly.
## Example
User: "Give me a 1-3-1 for adding retry logic to our API client."
Agent:
**Problem:** API calls to external services fail intermittently under load and need graceful retry handling.
**Options:**
- **Option A: Exponential backoff with jitter in a shared utility.**
Pros: Simple, well-understood pattern; single implementation shared across all callers.
Cons: No circuit-breaking; callers still block during retries.
- **Option B: Retry library (tenacity) as a decorator on each endpoint call.**
Pros: Configurable per-endpoint; built-in support for circuit breakers and rate limits.
Cons: Adds a dependency; decorator pattern can obscure control flow.
- **Option C: Queue-based async retry with dead letter handling.**
Pros: Non-blocking; failed calls are tracked and replayable; scales well.
Cons: Significant complexity increase; requires a queue backend; overkill for low-volume services.
**Recommendation:** Option A. The call volume doesn't justify queue infrastructure, and a shared utility with exponential backoff + jitter covers the failure mode cleanly without new dependencies.
**Definition of Done:**
- `api_retry()` utility handles transient HTTP errors (429, 502, 503, 504).
- Retries up to 3 times with exponential backoff and jitter.
- All external API calls route through the retry utility.
- Logs each retry attempt with endpoint and attempt number.
- Tests cover retry exhaustion, immediate success, and transient-then-success scenarios.
**Implementation Plan:**
1. Create `utils/api_retry.py` with configurable max retries, base delay, and retryable status codes.
2. Add jitter using `random.uniform(0, base_delay)` to prevent thundering herd.
3. Wrap existing API calls in `api_client.py` with the retry utility.
4. Add unit tests mocking HTTP responses for each retry scenario.
5. Verify under load with a simple stress test against a flaky endpoint mock.

134
website/docs/user-guide/skills/optional/creative/creative-blender-mcp.md Normal file
View file

@ -0,0 +1,134 @@
---
title: "Blender Mcp — Control Blender directly from Hermes via socket connection to the blender-mcp addon"
sidebar_label: "Blender Mcp"
description: "Control Blender directly from Hermes via socket connection to the blender-mcp addon"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Blender Mcp
Control Blender directly from Hermes via socket connection to the blender-mcp addon. Create 3D objects, materials, animations, and run arbitrary Blender Python (bpy) code. Use when user wants to create or modify anything in Blender.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/creative/blender-mcp` |
| Path | `optional-skills/creative/blender-mcp` |
| Version | `1.0.0` |
| Author | alireza78a |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Blender MCP
Control a running Blender instance from Hermes via socket on TCP port 9876.
## Setup (one-time)
### 1. Install the Blender addon
curl -sL https://raw.githubusercontent.com/ahujasid/blender-mcp/main/addon.py -o ~/Desktop/blender_mcp_addon.py
In Blender:
Edit > Preferences > Add-ons > Install > select blender_mcp_addon.py
Enable "Interface: Blender MCP"
### 2. Start the socket server in Blender
Press N in Blender viewport to open sidebar.
Find "BlenderMCP" tab and click "Start Server".
### 3. Verify connection
nc -z -w2 localhost 9876 && echo "OPEN" || echo "CLOSED"
## Protocol
Plain UTF-8 JSON over TCP -- no length prefix.
Send: &#123;"type": "&lt;command>", "params": &#123;&lt;kwargs>&#125;&#125;
Receive: &#123;"status": "success", "result": &lt;value>&#125;
&#123;"status": "error", "message": "&lt;reason>"&#125;
## Available Commands
| type | params | description |
|-------------------------|-------------------|---------------------------------|
| execute_code | code (str) | Run arbitrary bpy Python code |
| get_scene_info | (none) | List all objects in scene |
| get_object_info | object_name (str) | Details on a specific object |
| get_viewport_screenshot | (none) | Screenshot of current viewport |
## Python Helper
Use this inside execute_code tool calls:
import socket, json
def blender_exec(code: str, host="localhost", port=9876, timeout=15):
s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
s.connect((host, port))
s.settimeout(timeout)
payload = json.dumps(&#123;"type": "execute_code", "params": &#123;"code": code&#125;&#125;)
s.sendall(payload.encode("utf-8"))
buf = b""
while True:
try:
chunk = s.recv(4096)
if not chunk:
break
buf += chunk
try:
json.loads(buf.decode("utf-8"))
break
except json.JSONDecodeError:
continue
except socket.timeout:
break
s.close()
return json.loads(buf.decode("utf-8"))
## Common bpy Patterns
### Clear scene
bpy.ops.object.select_all(action='SELECT')
bpy.ops.object.delete()
### Add mesh objects
bpy.ops.mesh.primitive_uv_sphere_add(radius=1, location=(0, 0, 0))
bpy.ops.mesh.primitive_cube_add(size=2, location=(3, 0, 0))
bpy.ops.mesh.primitive_cylinder_add(radius=0.5, depth=2, location=(-3, 0, 0))
### Create and assign material
mat = bpy.data.materials.new(name="MyMat")
mat.use_nodes = True
bsdf = mat.node_tree.nodes.get("Principled BSDF")
bsdf.inputs["Base Color"].default_value = (R, G, B, 1.0)
bsdf.inputs["Roughness"].default_value = 0.3
bsdf.inputs["Metallic"].default_value = 0.0
obj.data.materials.append(mat)
### Keyframe animation
obj.location = (0, 0, 0)
obj.keyframe_insert(data_path="location", frame=1)
obj.location = (0, 0, 3)
obj.keyframe_insert(data_path="location", frame=60)
### Render to file
bpy.context.scene.render.filepath = "/tmp/render.png"
bpy.context.scene.render.engine = 'CYCLES'
bpy.ops.render.render(write_still=True)
## Pitfalls
- Must check socket is open before running (nc -z localhost 9876)
- Addon server must be started inside Blender each session (N-panel > BlenderMCP > Connect)
- Break complex scenes into multiple smaller execute_code calls to avoid timeouts
- Render output path must be absolute (/tmp/...) not relative
- shade_smooth() requires object to be selected and in object mode

378
website/docs/user-guide/skills/optional/creative/creative-concept-diagrams.md Normal file
View file

@ -0,0 +1,378 @@
---
title: "Concept Diagrams"
sidebar_label: "Concept Diagrams"
description: "Generate flat, minimal light/dark-aware SVG diagrams as standalone HTML files, using a unified educational visual language with 9 semantic color ramps, sente..."
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Concept Diagrams
Generate flat, minimal light/dark-aware SVG diagrams as standalone HTML files, using a unified educational visual language with 9 semantic color ramps, sentence-case typography, and automatic dark mode. Best suited for educational and non-software visuals — physics setups, chemistry mechanisms, math curves, physical objects (aircraft, turbines, smartphones, mechanical watches), anatomy, floor plans, cross-sections, narrative journeys (lifecycle of X, process of Y), hub-spoke system integrations (smart city, IoT), and exploded layer views. If a more specialized skill exists for the subject (dedicated software/cloud architecture, hand-drawn sketches, animated explainers, etc.), prefer that — otherwise this skill can also serve as a general-purpose SVG diagram fallback with a clean educational look. Ships with 15 example diagrams.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/creative/concept-diagrams` |
| Path | `optional-skills/creative/concept-diagrams` |
| Version | `0.1.0` |
| Author | v1k22 (original PR), ported into hermes-agent |
| License | MIT |
| Tags | `diagrams`, `svg`, `visualization`, `education`, `physics`, `chemistry`, `engineering` |
| Related skills | [`architecture-diagram`](/docs/user-guide/skills/bundled/creative/creative-architecture-diagram), [`excalidraw`](/docs/user-guide/skills/bundled/creative/creative-excalidraw), `generative-widgets` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Concept Diagrams
Generate production-quality SVG diagrams with a unified flat, minimal design system. Output is a single self-contained HTML file that renders identically in any modern browser, with automatic light/dark mode.
## Scope
**Best suited for:**
- Physics setups, chemistry mechanisms, math curves, biology
- Physical objects (aircraft, turbines, smartphones, mechanical watches, cells)
- Anatomy, cross-sections, exploded layer views
- Floor plans, architectural conversions
- Narrative journeys (lifecycle of X, process of Y)
- Hub-spoke system integrations (smart city, IoT networks, electricity grids)
- Educational / textbook-style visuals in any domain
- Quantitative charts (grouped bars, energy profiles)
**Look elsewhere first for:**
- Dedicated software / cloud infrastructure architecture with a dark tech aesthetic (consider `architecture-diagram` if available)
- Hand-drawn whiteboard sketches (consider `excalidraw` if available)
- Animated explainers or video output (consider an animation skill)
If a more specialized skill is available for the subject, prefer that. If none fits, this skill can serve as a general-purpose SVG diagram fallback — the output will carry the clean educational aesthetic described below, which is a reasonable default for almost any subject.
## Workflow
1. Decide on the diagram type (see Diagram Types below).
2. Lay out components using the Design System rules.
3. Write the full HTML page using `templates/template.html` as the wrapper — paste your SVG where the template says `<!-- PASTE SVG HERE -->`.
4. Save as a standalone `.html` file (for example `~/my-diagram.html` or `./my-diagram.html`).
5. User opens it directly in a browser — no server, no dependencies.
Optional: if the user wants a browsable gallery of multiple diagrams, see "Local Preview Server" at the bottom.
Load the HTML template:
```
skill_view(name="concept-diagrams", file_path="templates/template.html")
```
The template embeds the full CSS design system (`c-*` color classes, text classes, light/dark variables, arrow marker styles). The SVG you generate relies on these classes being present on the hosting page.
---
## Design System
### Philosophy
- **Flat**: no gradients, drop shadows, blur, glow, or neon effects.
- **Minimal**: show the essential. No decorative icons inside boxes.
- **Consistent**: same colors, spacing, typography, and stroke widths across every diagram.
- **Dark-mode ready**: all colors auto-adapt via CSS classes — no per-mode SVG.
### Color Palette
9 color ramps, each with 7 stops. Put the class name on a `<g>` or shape element; the template CSS handles both modes.
| Class | 50 (lightest) | 100 | 200 | 400 | 600 | 800 | 900 (darkest) |
|------------|---------------|---------|---------|---------|---------|---------|---------------|
| `c-purple` | #EEEDFE | #CECBF6 | #AFA9EC | #7F77DD | #534AB7 | #3C3489 | #26215C |
| `c-teal` | #E1F5EE | #9FE1CB | #5DCAA5 | #1D9E75 | #0F6E56 | #085041 | #04342C |
| `c-coral` | #FAECE7 | #F5C4B3 | #F0997B | #D85A30 | #993C1D | #712B13 | #4A1B0C |
| `c-pink` | #FBEAF0 | #F4C0D1 | #ED93B1 | #D4537E | #993556 | #72243E | #4B1528 |
| `c-gray` | #F1EFE8 | #D3D1C7 | #B4B2A9 | #888780 | #5F5E5A | #444441 | #2C2C2A |
| `c-blue` | #E6F1FB | #B5D4F4 | #85B7EB | #378ADD | #185FA5 | #0C447C | #042C53 |
| `c-green` | #EAF3DE | #C0DD97 | #97C459 | #639922 | #3B6D11 | #27500A | #173404 |
| `c-amber` | #FAEEDA | #FAC775 | #EF9F27 | #BA7517 | #854F0B | #633806 | #412402 |
| `c-red` | #FCEBEB | #F7C1C1 | #F09595 | #E24B4A | #A32D2D | #791F1F | #501313 |
#### Color Assignment Rules
Color encodes **meaning**, not sequence. Never cycle through colors like a rainbow.
- Group nodes by **category** — all nodes of the same type share one color.
- Use `c-gray` for neutral/structural nodes (start, end, generic steps, users).
- Use **2-3 colors per diagram**, not 6+.
- Prefer `c-purple`, `c-teal`, `c-coral`, `c-pink` for general categories.
- Reserve `c-blue`, `c-green`, `c-amber`, `c-red` for semantic meaning (info, success, warning, error).
Light/dark stop mapping (handled by the template CSS — just use the class):
- Light mode: 50 fill + 600 stroke + 800 title / 600 subtitle
- Dark mode: 800 fill + 200 stroke + 100 title / 200 subtitle
### Typography
Only two font sizes. No exceptions.
| Class | Size | Weight | Use |
|-------|------|--------|-----|
| `th` | 14px | 500 | Node titles, region labels |
| `ts` | 12px | 400 | Subtitles, descriptions, arrow labels |
| `t` | 14px | 400 | General text |
- **Sentence case always.** Never Title Case, never ALL CAPS.
- Every `<text>` MUST carry a class (`t`, `ts`, or `th`). No unclassed text.
- `dominant-baseline="central"` on all text inside boxes.
- `text-anchor="middle"` for centered text in boxes.
**Width estimation (approx):**
- 14px weight 500: ~8px per character
- 12px weight 400: ~6.5px per character
- Always verify: `box_width >= (char_count × px_per_char) + 48` (24px padding each side)
### Spacing & Layout
- **ViewBox**: `viewBox="0 0 680 H"` where H = content height + 40px buffer.
- **Safe area**: x=40 to x=640, y=40 to y=(H-40).
- **Between boxes**: 60px minimum gap.
- **Inside boxes**: 24px horizontal padding, 12px vertical padding.
- **Arrowhead gap**: 10px between arrowhead and box edge.
- **Single-line box**: 44px height.
- **Two-line box**: 56px height, 18px between title and subtitle baselines.
- **Container padding**: 20px minimum inside every container.
- **Max nesting**: 2-3 levels deep. Deeper gets unreadable at 680px width.
### Stroke & Shape
- **Stroke width**: 0.5px on all node borders. Not 1px, not 2px.
- **Rect rounding**: `rx="8"` for nodes, `rx="12"` for inner containers, `rx="16"` to `rx="20"` for outer containers.
- **Connector paths**: MUST have `fill="none"`. SVG defaults to `fill: black` otherwise.
### Arrow Marker
Include this `<defs>` block at the start of **every** SVG:
```xml
<defs>
<marker id="arrow" viewBox="0 0 10 10" refX="8" refY="5"
markerWidth="6" markerHeight="6" orient="auto-start-reverse">
<path d="M2 1L8 5L2 9" fill="none" stroke="context-stroke"
stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round"/>
</marker>
</defs>
```
Use `marker-end="url(#arrow)"` on lines. The arrowhead inherits the line color via `context-stroke`.
### CSS Classes (Provided by the Template)
The template page provides:
- Text: `.t`, `.ts`, `.th`
- Neutral: `.box`, `.arr`, `.leader`, `.node`
- Color ramps: `.c-purple`, `.c-teal`, `.c-coral`, `.c-pink`, `.c-gray`, `.c-blue`, `.c-green`, `.c-amber`, `.c-red` (all with automatic light/dark mode)
You do **not** need to redefine these — just apply them in your SVG. The template file contains the full CSS definitions.
---
## SVG Boilerplate
Every SVG inside the template page starts with this exact structure:
```xml
<svg width="100%" viewBox="0 0 680 {HEIGHT}" xmlns="http://www.w3.org/2000/svg">
<defs>
<marker id="arrow" viewBox="0 0 10 10" refX="8" refY="5"
markerWidth="6" markerHeight="6" orient="auto-start-reverse">
<path d="M2 1L8 5L2 9" fill="none" stroke="context-stroke"
stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round"/>
</marker>
</defs>
<!-- Diagram content here -->
</svg>
```
Replace `{HEIGHT}` with the actual computed height (last element bottom + 40px).
### Node Patterns
**Single-line node (44px):**
```xml
<g class="node c-blue">
<rect x="100" y="20" width="180" height="44" rx="8" stroke-width="0.5"/>
<text class="th" x="190" y="42" text-anchor="middle" dominant-baseline="central">Service name</text>
</g>
```
**Two-line node (56px):**
```xml
<g class="node c-teal">
<rect x="100" y="20" width="200" height="56" rx="8" stroke-width="0.5"/>
<text class="th" x="200" y="38" text-anchor="middle" dominant-baseline="central">Service name</text>
<text class="ts" x="200" y="56" text-anchor="middle" dominant-baseline="central">Short description</text>
</g>
```
**Connector (no label):**
```xml
<line x1="200" y1="76" x2="200" y2="120" class="arr" marker-end="url(#arrow)"/>
```
**Container (dashed or solid):**
```xml
<g class="c-purple">
<rect x="40" y="92" width="600" height="300" rx="16" stroke-width="0.5"/>
<text class="th" x="66" y="116">Container label</text>
<text class="ts" x="66" y="134">Subtitle info</text>
</g>
```
---
## Diagram Types
Choose the layout that fits the subject:
1. **Flowchart** — CI/CD pipelines, request lifecycles, approval workflows, data processing. Single-direction flow (top-down or left-right). Max 4-5 nodes per row.
2. **Structural / Containment** — Cloud infrastructure nesting, system architecture with layers. Large outer containers with inner regions. Dashed rects for logical groupings.
3. **API / Endpoint Map** — REST routes, GraphQL schemas. Tree from root, branching to resource groups, each containing endpoint nodes.
4. **Microservice Topology** — Service mesh, event-driven systems. Services as nodes, arrows for communication patterns, message queues between.
5. **Data Flow** — ETL pipelines, streaming architectures. Left-to-right flow from sources through processing to sinks.
6. **Physical / Structural** — Vehicles, buildings, hardware, anatomy. Use shapes that match the physical form — `<path>` for curved bodies, `<polygon>` for tapered shapes, `<ellipse>`/`<circle>` for cylindrical parts, nested `<rect>` for compartments. See `references/physical-shape-cookbook.md`.
7. **Infrastructure / Systems Integration** — Smart cities, IoT networks, multi-domain systems. Hub-spoke layout with central platform connecting subsystems. Semantic line styles (`.data-line`, `.power-line`, `.water-pipe`, `.road`). See `references/infrastructure-patterns.md`.
8. **UI / Dashboard Mockups** — Admin panels, monitoring dashboards. Screen frame with nested chart/gauge/indicator elements. See `references/dashboard-patterns.md`.
For physical, infrastructure, and dashboard diagrams, load the matching reference file before generating — each one provides ready-made CSS classes and shape primitives.
---
## Validation Checklist
Before finalizing any SVG, verify ALL of the following:
1. Every `<text>` has class `t`, `ts`, or `th`.
2. Every `<text>` inside a box has `dominant-baseline="central"`.
3. Every connector `<path>` or `<line>` used as arrow has `fill="none"`.
4. No arrow line crosses through an unrelated box.
5. `box_width >= (longest_label_chars × 8) + 48` for 14px text.
6. `box_width >= (longest_label_chars × 6.5) + 48` for 12px text.
7. ViewBox height = bottom-most element + 40px.
8. All content stays within x=40 to x=640.
9. Color classes (`c-*`) are on `<g>` or shape elements, never on `<path>` connectors.
10. Arrow `<defs>` block is present.
11. No gradients, shadows, blur, or glow effects.
12. Stroke width is 0.5px on all node borders.
---
## Output & Preview
### Default: standalone HTML file
Write a single `.html` file the user can open directly. No server, no dependencies, works offline. Pattern:
```python
# 1. Load the template
template = skill_view("concept-diagrams", "templates/template.html")
# 2. Fill in title, subtitle, and paste your SVG
html = template.replace(
"<!-- DIAGRAM TITLE HERE -->", "SN2 reaction mechanism"
).replace(
"<!-- OPTIONAL SUBTITLE HERE -->", "Bimolecular nucleophilic substitution"
).replace(
"<!-- PASTE SVG HERE -->", svg_content
)
# 3. Write to a user-chosen path (or ./ by default)
write_file("./sn2-mechanism.html", html)
```
Tell the user how to open it:
```
# macOS
open ./sn2-mechanism.html
# Linux
xdg-open ./sn2-mechanism.html
```
### Optional: local preview server (multi-diagram gallery)
Only use this when the user explicitly wants a browsable gallery of multiple diagrams.
**Rules:**
- Bind to `127.0.0.1` only. Never `0.0.0.0`. Exposing diagrams on all network interfaces is a security hazard on shared networks.
- Pick a free port (do NOT hard-code one) and tell the user the chosen URL.
- The server is optional and opt-in — prefer the standalone HTML file first.
Recommended pattern (lets the OS pick a free ephemeral port):
```bash
# Put each diagram in its own folder under .diagrams/
mkdir -p .diagrams/sn2-mechanism
# ...write .diagrams/sn2-mechanism/index.html...
# Serve on loopback only, free port
cd .diagrams && python3 -c "
import http.server, socketserver
with socketserver.TCPServer(('127.0.0.1', 0), http.server.SimpleHTTPRequestHandler) as s:
print(f'Serving at http://127.0.0.1:{s.server_address[1]}/')
s.serve_forever()
" &
```
If the user insists on a fixed port, use `127.0.0.1:<port>` — still never `0.0.0.0`. Document how to stop the server (`kill %1` or `pkill -f "http.server"`).
---
## Examples Reference
The `examples/` directory ships 15 complete, tested diagrams. Browse them for working patterns before writing a new diagram of a similar type:
| File | Type | Demonstrates |
|------|------|--------------|
| `hospital-emergency-department-flow.md` | Flowchart | Priority routing with semantic colors |
| `feature-film-production-pipeline.md` | Flowchart | Phased workflow, horizontal sub-flows |
| `automated-password-reset-flow.md` | Flowchart | Auth flow with error branches |
| `autonomous-llm-research-agent-flow.md` | Flowchart | Loop-back arrows, decision branches |
| `place-order-uml-sequence.md` | Sequence | UML sequence diagram style |
| `commercial-aircraft-structure.md` | Physical | Paths, polygons, ellipses for realistic shapes |
| `wind-turbine-structure.md` | Physical cross-section | Underground/above-ground separation, color coding |
| `smartphone-layer-anatomy.md` | Exploded view | Alternating left/right labels, layered components |
| `apartment-floor-plan-conversion.md` | Floor plan | Walls, doors, proposed changes in dotted red |
| `banana-journey-tree-to-smoothie.md` | Narrative journey | Winding path, progressive state changes |
| `cpu-ooo-microarchitecture.md` | Hardware pipeline | Fan-out, memory hierarchy sidebar |
| `sn2-reaction-mechanism.md` | Chemistry | Molecules, curved arrows, energy profile |
| `smart-city-infrastructure.md` | Hub-spoke | Semantic line styles per system |
| `electricity-grid-flow.md` | Multi-stage flow | Voltage hierarchy, flow markers |
| `ml-benchmark-grouped-bar-chart.md` | Chart | Grouped bars, dual axis |
Load any example with:
```
skill_view(name="concept-diagrams", file_path="examples/<filename>")
```
---
## Quick Reference: What to Use When
| User says | Diagram type | Suggested colors |
|-----------|--------------|------------------|
| "show the pipeline" | Flowchart | gray start/end, purple steps, red errors, teal deploy |
| "draw the data flow" | Data pipeline (left-right) | gray sources, purple processing, teal sinks |
| "visualize the system" | Structural (containment) | purple container, teal services, coral data |
| "map the endpoints" | API tree | purple root, one ramp per resource group |
| "show the services" | Microservice topology | gray ingress, teal services, purple bus, coral workers |
| "draw the aircraft/vehicle" | Physical | paths, polygons, ellipses for realistic shapes |
| "smart city / IoT" | Hub-spoke integration | semantic line styles per subsystem |
| "show the dashboard" | UI mockup | dark screen, chart colors: teal, purple, coral for alerts |
| "power grid / electricity" | Multi-stage flow | voltage hierarchy (HV/MV/LV line weights) |
| "wind turbine / turbine" | Physical cross-section | foundation + tower cutaway + nacelle color-coded |
| "journey of X / lifecycle" | Narrative journey | winding path, progressive state changes |
| "layers of X / exploded" | Exploded layer view | vertical stack, alternating labels |
| "CPU / pipeline" | Hardware pipeline | vertical stages, fan-out to execution ports |
| "floor plan / apartment" | Floor plan | walls, doors, proposed changes in dotted red |
| "reaction mechanism" | Chemistry | atoms, bonds, curved arrows, transition state, energy profile |

146
website/docs/user-guide/skills/optional/creative/creative-meme-generation.md Normal file
View file

@ -0,0 +1,146 @@
---
title: "Meme Generation — Generate real meme images by picking a template and overlaying text with Pillow"
sidebar_label: "Meme Generation"
description: "Generate real meme images by picking a template and overlaying text with Pillow"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Meme Generation
Generate real meme images by picking a template and overlaying text with Pillow. Produces actual .png meme files.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/creative/meme-generation` |
| Path | `optional-skills/creative/meme-generation` |
| Version | `2.0.0` |
| Author | adanaleycio |
| License | MIT |
| Tags | `creative`, `memes`, `humor`, `images` |
| Related skills | [`ascii-art`](/docs/user-guide/skills/bundled/creative/creative-ascii-art), `generative-widgets` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Meme Generation
Generate actual meme images from a topic. Picks a template, writes captions, and renders a real .png file with text overlay.
## When to Use
- User asks you to make or generate a meme
- User wants a meme about a specific topic, situation, or frustration
- User says "meme this" or similar
## Available Templates
The script supports **any of the ~100 popular imgflip templates** by name or ID, plus 10 curated templates with hand-tuned text positioning.
### Curated Templates (custom text placement)
| ID | Name | Fields | Best for |
|----|------|--------|----------|
| `this-is-fine` | This is Fine | top, bottom | chaos, denial |
| `drake` | Drake Hotline Bling | reject, approve | rejecting/preferring |
| `distracted-boyfriend` | Distracted Boyfriend | distraction, current, person | temptation, shifting priorities |
| `two-buttons` | Two Buttons | left, right, person | impossible choice |
| `expanding-brain` | Expanding Brain | 4 levels | escalating irony |
| `change-my-mind` | Change My Mind | statement | hot takes |
| `woman-yelling-at-cat` | Woman Yelling at Cat | woman, cat | arguments |
| `one-does-not-simply` | One Does Not Simply | top, bottom | deceptively hard things |
| `grus-plan` | Gru's Plan | step1-3, realization | plans that backfire |
| `batman-slapping-robin` | Batman Slapping Robin | robin, batman | shutting down bad ideas |
### Dynamic Templates (from imgflip API)
Any template not in the curated list can be used by name or imgflip ID. These get smart default text positioning (top/bottom for 2-field, evenly spaced for 3+). Search with:
```bash
python "$SKILL_DIR/scripts/generate_meme.py" --search "disaster"
```
## Procedure
### Mode 1: Classic Template (default)
1. Read the user's topic and identify the core dynamic (chaos, dilemma, preference, irony, etc.)
2. Pick the template that best matches. Use the "Best for" column, or search with `--search`.
3. Write short captions for each field (8-12 words max per field, shorter is better).
4. Find the skill's script directory:
```
SKILL_DIR=$(dirname "$(find ~/.hermes/skills -path '*/meme-generation/SKILL.md' 2>/dev/null | head -1)")
```
5. Run the generator:
```bash
python "$SKILL_DIR/scripts/generate_meme.py" <template_id> /tmp/meme.png "caption 1" "caption 2" ...
```
6. Return the image with `MEDIA:/tmp/meme.png`
### Mode 2: Custom AI Image (when image_generate is available)
Use this when no classic template fits, or when the user wants something original.
1. Write the captions first.
2. Use `image_generate` to create a scene that matches the meme concept. Do NOT include any text in the image prompt — text will be added by the script. Describe only the visual scene.
3. Find the generated image path from the image_generate result URL. Download it to a local path if needed.
4. Run the script with `--image` to overlay text, choosing a mode:
- **Overlay** (text directly on image, white with black outline):
```bash
python "$SKILL_DIR/scripts/generate_meme.py" --image /path/to/scene.png /tmp/meme.png "top text" "bottom text"
```
- **Bars** (black bars above/below with white text — cleaner, always readable):
```bash
python "$SKILL_DIR/scripts/generate_meme.py" --image /path/to/scene.png --bars /tmp/meme.png "top text" "bottom text"
```
Use `--bars` when the image is busy/detailed and text would be hard to read on top of it.
5. **Verify with vision** (if `vision_analyze` is available): Check the result looks good:
```
vision_analyze(image_url="/tmp/meme.png", question="Is the text legible and well-positioned? Does the meme work visually?")
```
If the vision model flags issues (text hard to read, bad placement, etc.), try the other mode (switch between overlay and bars) or regenerate the scene.
6. Return the image with `MEDIA:/tmp/meme.png`
## Examples
**"debugging production at 2 AM":**
```bash
python generate_meme.py this-is-fine /tmp/meme.png "SERVERS ARE ON FIRE" "This is fine"
```
**"choosing between sleep and one more episode":**
```bash
python generate_meme.py drake /tmp/meme.png "Getting 8 hours of sleep" "One more episode at 3 AM"
```
**"the stages of a Monday morning":**
```bash
python generate_meme.py expanding-brain /tmp/meme.png "Setting an alarm" "Setting 5 alarms" "Sleeping through all alarms" "Working from bed"
```
## Listing Templates
To see all available templates:
```bash
python generate_meme.py --list
```
## Pitfalls
- Keep captions SHORT. Memes with long text look terrible.
- Match the number of text arguments to the template's field count.
- Pick the template that fits the joke structure, not just the topic.
- Do not generate hateful, abusive, or personally targeted content.
- The script caches template images in `scripts/.cache/` after first download.
## Verification
The output is correct if:
- A .png file was created at the output path
- Text is legible (white with black outline) on the template
- The joke lands — caption matches the template's intended structure
- File can be delivered via MEDIA: path

356
website/docs/user-guide/skills/optional/creative/creative-touchdesigner-mcp.md Normal file
View file

@ -0,0 +1,356 @@
---
title: "Touchdesigner Mcp"
sidebar_label: "Touchdesigner Mcp"
description: "Control a running TouchDesigner instance via twozero MCP — create operators, set parameters, wire connections, execute Python, build real-time visuals"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Touchdesigner Mcp
Control a running TouchDesigner instance via twozero MCP — create operators, set parameters, wire connections, execute Python, build real-time visuals. 36 native tools.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/creative/touchdesigner-mcp` |
| Path | `optional-skills/creative/touchdesigner-mcp` |
| Version | `1.0.0` |
| Author | kshitijk4poor |
| License | MIT |
| Tags | `TouchDesigner`, `MCP`, `twozero`, `creative-coding`, `real-time-visuals`, `generative-art`, `audio-reactive`, `VJ`, `installation`, `GLSL` |
| Related skills | [`native-mcp`](/docs/user-guide/skills/bundled/mcp/mcp-native-mcp), [`ascii-video`](/docs/user-guide/skills/bundled/creative/creative-ascii-video), [`manim-video`](/docs/user-guide/skills/bundled/creative/creative-manim-video), `hermes-video` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# TouchDesigner Integration (twozero MCP)
## CRITICAL RULES
1. **NEVER guess parameter names.** Call `td_get_par_info` for the op type FIRST. Your training data is wrong for TD 2025.32.
2. **If `tdAttributeError` fires, STOP.** Call `td_get_operator_info` on the failing node before continuing.
3. **NEVER hardcode absolute paths** in script callbacks. Use `me.parent()` / `scriptOp.parent()`.
4. **Prefer native MCP tools over td_execute_python.** Use `td_create_operator`, `td_set_operator_pars`, `td_get_errors` etc. Only fall back to `td_execute_python` for complex multi-step logic.
5. **Call `td_get_hints` before building.** It returns patterns specific to the op type you're working with.
## Architecture
```
Hermes Agent -> MCP (Streamable HTTP) -> twozero.tox (port 40404) -> TD Python
```
36 native tools. Free plugin (no payment/license — confirmed April 2026).
Context-aware (knows selected OP, current network).
Hub health check: `GET http://localhost:40404/mcp` returns JSON with instance PID, project name, TD version.
## Setup (Automated)
Run the setup script to handle everything:
```bash
bash "${HERMES_HOME:-$HOME/.hermes}/skills/creative/touchdesigner-mcp/scripts/setup.sh"
```
The script will:
1. Check if TD is running
2. Download twozero.tox if not already cached
3. Add `twozero_td` MCP server to Hermes config (if missing)
4. Test the MCP connection on port 40404
5. Report what manual steps remain (drag .tox into TD, enable MCP toggle)
### Manual steps (one-time, cannot be automated)
1. **Drag `~/Downloads/twozero.tox` into the TD network editor** → click Install
2. **Enable MCP:** click twozero icon → Settings → mcp → "auto start MCP" → Yes
3. **Restart Hermes session** to pick up the new MCP server
After setup, verify:
```bash
nc -z 127.0.0.1 40404 && echo "twozero MCP: READY"
```
## Environment Notes
- **Non-Commercial TD** caps resolution at 1280×1280. Use `outputresolution = 'custom'` and set width/height explicitly.
- **Codecs:** `prores` (preferred on macOS) or `mjpa` as fallback. H.264/H.265/AV1 require a Commercial license.
- Always call `td_get_par_info` before setting params — names vary by TD version (see CRITICAL RULES #1).
## Workflow
### Step 0: Discover (before building anything)
```
Call td_get_par_info with op_type for each type you plan to use.
Call td_get_hints with the topic you're building (e.g. "glsl", "audio reactive", "feedback").
Call td_get_focus to see where the user is and what's selected.
Call td_get_network to see what already exists.
```
No temp nodes, no cleanup. This replaces the old discovery dance entirely.
### Step 1: Clean + Build
**IMPORTANT: Split cleanup and creation into SEPARATE MCP calls.** Destroying and recreating same-named nodes in one `td_execute_python` script causes "Invalid OP object" errors. See pitfalls #11b.
Use `td_create_operator` for each node (handles viewport positioning automatically):
```
td_create_operator(type="noiseTOP", parent="/project1", name="bg", parameters={"resolutionw": 1280, "resolutionh": 720})
td_create_operator(type="levelTOP", parent="/project1", name="brightness")
td_create_operator(type="nullTOP", parent="/project1", name="out")
```
For bulk creation or wiring, use `td_execute_python`:
```python
# td_execute_python script:
root = op('/project1')
nodes = []
for name, optype in [('bg', noiseTOP), ('fx', levelTOP), ('out', nullTOP)]:
n = root.create(optype, name)
nodes.append(n.path)
# Wire chain
for i in range(len(nodes)-1):
op(nodes[i]).outputConnectors[0].connect(op(nodes[i+1]).inputConnectors[0])
result = {'created': nodes}
```
### Step 2: Set Parameters
Prefer the native tool (validates params, won't crash):
```
td_set_operator_pars(path="/project1/bg", parameters={"roughness": 0.6, "monochrome": true})
```
For expressions or modes, use `td_execute_python`:
```python
op('/project1/time_driver').par.colorr.expr = "absTime.seconds % 1000.0"
```
### Step 3: Wire
Use `td_execute_python` — no native wire tool exists:
```python
op('/project1/bg').outputConnectors[0].connect(op('/project1/fx').inputConnectors[0])
```
### Step 4: Verify
```
td_get_errors(path="/project1", recursive=true)
td_get_perf()
td_get_operator_info(path="/project1/out", detail="full")
```
### Step 5: Display / Capture
```
td_get_screenshot(path="/project1/out")
```
Or open a window via script:
```python
win = op('/project1').create(windowCOMP, 'display')
win.par.winop = op('/project1/out').path
win.par.winw = 1280; win.par.winh = 720
win.par.winopen.pulse()
```
## MCP Tool Quick Reference
**Core (use these most):**
| Tool | What |
|------|------|
| `td_execute_python` | Run arbitrary Python in TD. Full API access. |
| `td_create_operator` | Create node with params + auto-positioning |
| `td_set_operator_pars` | Set params safely (validates, won't crash) |
| `td_get_operator_info` | Inspect one node: connections, params, errors |
| `td_get_operators_info` | Inspect multiple nodes in one call |
| `td_get_network` | See network structure at a path |
| `td_get_errors` | Find errors/warnings recursively |
| `td_get_par_info` | Get param names for an OP type (replaces discovery) |
| `td_get_hints` | Get patterns/tips before building |
| `td_get_focus` | What network is open, what's selected |
**Read/Write:**
| Tool | What |
|------|------|
| `td_read_dat` | Read DAT text content |
| `td_write_dat` | Write/patch DAT content |
| `td_read_chop` | Read CHOP channel values |
| `td_read_textport` | Read TD console output |
**Visual:**
| Tool | What |
|------|------|
| `td_get_screenshot` | Capture one OP viewer to file |
| `td_get_screenshots` | Capture multiple OPs at once |
| `td_get_screen_screenshot` | Capture actual screen via TD |
| `td_navigate_to` | Jump network editor to an OP |
**Search:**
| Tool | What |
|------|------|
| `td_find_op` | Find ops by name/type across project |
| `td_search` | Search code, expressions, string params |
**System:**
| Tool | What |
|------|------|
| `td_get_perf` | Performance profiling (FPS, slow ops) |
| `td_list_instances` | List all running TD instances |
| `td_get_docs` | In-depth docs on a TD topic |
| `td_agents_md` | Read/write per-COMP markdown docs |
| `td_reinit_extension` | Reload extension after code edit |
| `td_clear_textport` | Clear console before debug session |
**Input Automation:**
| Tool | What |
|------|------|
| `td_input_execute` | Send mouse/keyboard to TD |
| `td_input_status` | Poll input queue status |
| `td_input_clear` | Stop input automation |
| `td_op_screen_rect` | Get screen coords of a node |
| `td_click_screen_point` | Click a point in a screenshot |
See `references/mcp-tools.md` for full parameter schemas.
## Key Implementation Rules
**GLSL time:** No `uTDCurrentTime` in GLSL TOP. Use the Values page:
```python
# Call td_get_par_info(op_type="glslTOP") first to confirm param names
td_set_operator_pars(path="/project1/shader", parameters={"value0name": "uTime"})
# Then set expression via script:
# op('/project1/shader').par.value0.expr = "absTime.seconds"
# In GLSL: uniform float uTime;
```
Fallback: Constant TOP in `rgba32float` format (8-bit clamps to 0-1, freezing the shader).
**Feedback TOP:** Use `top` parameter reference, not direct input wire. "Not enough sources" resolves after first cook. "Cook dependency loop" warning is expected.
**Resolution:** Non-Commercial caps at 1280×1280. Use `outputresolution = 'custom'`.
**Large shaders:** Write GLSL to `/tmp/file.glsl`, then use `td_write_dat` or `td_execute_python` to load.
**Vertex/Point access (TD 2025.32):** `point.P[0]`, `point.P[1]`, `point.P[2]` — NOT `.x`, `.y`, `.z`.
**Extensions:** `ext0object` format is `"op('./datName').module.ClassName(me)"` in CONSTANT mode. After editing extension code with `td_write_dat`, call `td_reinit_extension`.
**Script callbacks:** ALWAYS use relative paths via `me.parent()` / `scriptOp.parent()`.
**Cleaning nodes:** Always `list(root.children)` before iterating + `child.valid` check.
## Recording / Exporting Video
```python
# via td_execute_python:
root = op('/project1')
rec = root.create(moviefileoutTOP, 'recorder')
op('/project1/out').outputConnectors[0].connect(rec.inputConnectors[0])
rec.par.type = 'movie'
rec.par.file = '/tmp/output.mov'
rec.par.videocodec = 'prores' # Apple ProRes — NOT license-restricted on macOS
rec.par.record = True # start
# rec.par.record = False # stop (call separately later)
```
H.264/H.265/AV1 need Commercial license. Use `prores` on macOS or `mjpa` as fallback.
Extract frames: `ffmpeg -i /tmp/output.mov -vframes 120 /tmp/frames/frame_%06d.png`
**TOP.save() is useless for animation** — captures same GPU texture every time. Always use MovieFileOut.
### Before Recording: Checklist
1. **Verify FPS > 0** via `td_get_perf`. If FPS=0 the recording will be empty. See pitfalls #38-39.
2. **Verify shader output is not black** via `td_get_screenshot`. Black output = shader error or missing input. See pitfalls #8, #40.
3. **If recording with audio:** cue audio to start first, then delay recording by 3 frames. See pitfalls #19.
4. **Set output path before starting record** — setting both in the same script can race.
## Audio-Reactive GLSL (Proven Recipe)
### Correct signal chain (tested April 2026)
```
AudioFileIn CHOP (playmode=sequential)
→ AudioSpectrum CHOP (FFT=512, outputmenu=setmanually, outlength=256, timeslice=ON)
→ Math CHOP (gain=10)
→ CHOP to TOP (dataformat=r, layout=rowscropped)
→ GLSL TOP input 1 (spectrum texture, 256x2)
Constant TOP (rgba32float, time) → GLSL TOP input 0
GLSL TOP → Null TOP → MovieFileOut
```
### Critical audio-reactive rules (empirically verified)
1. **TimeSlice must stay ON** for AudioSpectrum. OFF = processes entire audio file → 24000+ samples → CHOP to TOP overflow.
2. **Set Output Length manually** to 256 via `outputmenu='setmanually'` and `outlength=256`. Default outputs 22050 samples.
3. **DO NOT use Lag CHOP for spectrum smoothing.** Lag CHOP operates in timeslice mode and expands 256 samples to 2400+, averaging all values to near-zero (~1e-06). The shader receives no usable data. This was the #1 audio sync failure in testing.
4. **DO NOT use Filter CHOP either** — same timeslice expansion problem with spectrum data.
5. **Smoothing belongs in the GLSL shader** if needed, via temporal lerp with a feedback texture: `mix(prevValue, newValue, 0.3)`. This gives frame-perfect sync with zero pipeline latency.
6. **CHOP to TOP dataformat = 'r'**, layout = 'rowscropped'. Spectrum output is 256x2 (stereo). Sample at y=0.25 for first channel.
7. **Math gain = 10** (not 5). Raw spectrum values are ~0.19 in bass range. Gain of 10 gives usable ~5.0 for the shader.
8. **No Resample CHOP needed.** Control output size via AudioSpectrum's `outlength` param directly.
### GLSL spectrum sampling
```glsl
// Input 0 = time (1x1 rgba32float), Input 1 = spectrum (256x2)
float iTime = texture(sTD2DInputs[0], vec2(0.5)).r;
// Sample multiple points per band and average for stability:
// NOTE: y=0.25 for first channel (stereo texture is 256x2, first row center is 0.25)
float bass = (texture(sTD2DInputs[1], vec2(0.02, 0.25)).r +
texture(sTD2DInputs[1], vec2(0.05, 0.25)).r) / 2.0;
float mid = (texture(sTD2DInputs[1], vec2(0.2, 0.25)).r +
texture(sTD2DInputs[1], vec2(0.35, 0.25)).r) / 2.0;
float hi = (texture(sTD2DInputs[1], vec2(0.6, 0.25)).r +
texture(sTD2DInputs[1], vec2(0.8, 0.25)).r) / 2.0;
```
See `references/network-patterns.md` for complete build scripts + shader code.
## Operator Quick Reference
| Family | Color | Python class / MCP type | Suffix |
|--------|-------|-------------|--------|
| TOP | Purple | noiseTOP, glslTOP, compositeTOP, levelTop, blurTOP, textTOP, nullTOP | TOP |
| CHOP | Green | audiofileinCHOP, audiospectrumCHOP, mathCHOP, lfoCHOP, constantCHOP | CHOP |
| SOP | Blue | gridSOP, sphereSOP, transformSOP, noiseSOP | SOP |
| DAT | White | textDAT, tableDAT, scriptDAT, webserverDAT | DAT |
| MAT | Yellow | phongMAT, pbrMAT, glslMAT, constMAT | MAT |
| COMP | Gray | geometryCOMP, containerCOMP, cameraCOMP, lightCOMP, windowCOMP | COMP |
## Security Notes
- MCP runs on localhost only (port 40404). No authentication — any local process can send commands.
- `td_execute_python` has unrestricted access to the TD Python environment and filesystem as the TD process user.
- `setup.sh` downloads twozero.tox from the official 404zero.com URL. Verify the download if concerned.
- The skill never sends data outside localhost. All MCP communication is local.
## References
| File | What |
|------|------|
| `references/pitfalls.md` | Hard-won lessons from real sessions |
| `references/operators.md` | All operator families with params and use cases |
| `references/network-patterns.md` | Recipes: audio-reactive, generative, GLSL, instancing |
| `references/mcp-tools.md` | Full twozero MCP tool parameter schemas |
| `references/python-api.md` | TD Python: op(), scripting, extensions |
| `references/troubleshooting.md` | Connection diagnostics, debugging |
| `scripts/setup.sh` | Automated setup script |
---
> You're not writing code. You're conducting light.

172
website/docs/user-guide/skills/optional/devops/devops-cli.md Normal file
View file

@ -0,0 +1,172 @@
---
title: "Inference Sh Cli — Run 150+ AI apps via inference"
sidebar_label: "Inference Sh Cli"
description: "Run 150+ AI apps via inference"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Inference Sh Cli
Run 150+ AI apps via inference.sh CLI (infsh) — image generation, video creation, LLMs, search, 3D, social automation. Uses the terminal tool. Triggers: inference.sh, infsh, ai apps, flux, veo, image generation, video generation, seedream, seedance, tavily
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/devops/cli` |
| Path | `optional-skills/devops/cli` |
| Version | `1.0.0` |
| Author | okaris |
| License | MIT |
| Tags | `AI`, `image-generation`, `video`, `LLM`, `search`, `inference`, `FLUX`, `Veo`, `Claude` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# inference.sh CLI
Run 150+ AI apps in the cloud with a simple CLI. No GPU required.
All commands use the **terminal tool** to run `infsh` commands.
## When to Use
- User asks to generate images (FLUX, Reve, Seedream, Grok, Gemini image)
- User asks to generate video (Veo, Wan, Seedance, OmniHuman)
- User asks about inference.sh or infsh
- User wants to run AI apps without managing individual provider APIs
- User asks for AI-powered search (Tavily, Exa)
- User needs avatar/lipsync generation
## Prerequisites
The `infsh` CLI must be installed and authenticated. Check with:
```bash
infsh me
```
If not installed:
```bash
curl -fsSL https://cli.inference.sh | sh
infsh login
```
See `references/authentication.md` for full setup details.
## Workflow
### 1. Always Search First
Never guess app names — always search to find the correct app ID:
```bash
infsh app list --search flux
infsh app list --search video
infsh app list --search image
```
### 2. Run an App
Use the exact app ID from the search results. Always use `--json` for machine-readable output:
```bash
infsh app run <app-id> --input '{"prompt": "your prompt here"}' --json
```
### 3. Parse the Output
The JSON output contains URLs to generated media. Present these to the user with `MEDIA:<url>` for inline display.
## Common Commands
### Image Generation
```bash
# Search for image apps
infsh app list --search image
# FLUX Dev with LoRA
infsh app run falai/flux-dev-lora --input '{"prompt": "sunset over mountains", "num_images": 1}' --json
# Gemini image generation
infsh app run google/gemini-2-5-flash-image --input '{"prompt": "futuristic city", "num_images": 1}' --json
# Seedream (ByteDance)
infsh app run bytedance/seedream-5-lite --input '{"prompt": "nature scene"}' --json
# Grok Imagine (xAI)
infsh app run xai/grok-imagine-image --input '{"prompt": "abstract art"}' --json
```
### Video Generation
```bash
# Search for video apps
infsh app list --search video
# Veo 3.1 (Google)
infsh app run google/veo-3-1-fast --input '{"prompt": "drone shot of coastline"}' --json
# Seedance (ByteDance)
infsh app run bytedance/seedance-1-5-pro --input '{"prompt": "dancing figure", "resolution": "1080p"}' --json
# Wan 2.5
infsh app run falai/wan-2-5 --input '{"prompt": "person walking through city"}' --json
```
### Local File Uploads
The CLI automatically uploads local files when you provide a path:
```bash
# Upscale a local image
infsh app run falai/topaz-image-upscaler --input '{"image": "/path/to/photo.jpg", "upscale_factor": 2}' --json
# Image-to-video from local file
infsh app run falai/wan-2-5-i2v --input '{"image": "/path/to/image.png", "prompt": "make it move"}' --json
# Avatar with audio
infsh app run bytedance/omnihuman-1-5 --input '{"audio": "/path/to/audio.mp3", "image": "/path/to/face.jpg"}' --json
```
### Search & Research
```bash
infsh app list --search search
infsh app run tavily/tavily-search --input '{"query": "latest AI news"}' --json
infsh app run exa/exa-search --input '{"query": "machine learning papers"}' --json
```
### Other Categories
```bash
# 3D generation
infsh app list --search 3d
# Audio / TTS
infsh app list --search tts
# Twitter/X automation
infsh app list --search twitter
```
## Pitfalls
1. **Never guess app IDs** — always run `infsh app list --search <term>` first. App IDs change and new apps are added frequently.
2. **Always use `--json`** — raw output is hard to parse. The `--json` flag gives structured output with URLs.
3. **Check authentication** — if commands fail with auth errors, run `infsh login` or verify `INFSH_API_KEY` is set.
4. **Long-running apps** — video generation can take 30-120 seconds. The terminal tool timeout should be sufficient, but warn the user it may take a moment.
5. **Input format** — the `--input` flag takes a JSON string. Make sure to properly escape quotes.
## Reference Docs
- `references/authentication.md` — Setup, login, API keys
- `references/app-discovery.md` — Searching and browsing the app catalog
- `references/running-apps.md` — Running apps, input formats, output handling
- `references/cli-reference.md` — Complete CLI command reference

296
website/docs/user-guide/skills/optional/devops/devops-docker-management.md Normal file
View file

@ -0,0 +1,296 @@
---
title: "Docker Management"
sidebar_label: "Docker Management"
description: "Manage Docker containers, images, volumes, networks, and Compose stacks — lifecycle ops, debugging, cleanup, and Dockerfile optimization"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Docker Management
Manage Docker containers, images, volumes, networks, and Compose stacks — lifecycle ops, debugging, cleanup, and Dockerfile optimization.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/devops/docker-management` |
| Path | `optional-skills/devops/docker-management` |
| Version | `1.0.0` |
| Author | sprmn24 |
| License | MIT |
| Tags | `docker`, `containers`, `devops`, `infrastructure`, `compose`, `images`, `volumes`, `networks`, `debugging` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Docker Management
Manage Docker containers, images, volumes, networks, and Compose stacks using standard Docker CLI commands. No additional dependencies beyond Docker itself.
## When to Use
- Run, stop, restart, remove, or inspect containers
- Build, pull, push, tag, or clean up Docker images
- Work with Docker Compose (multi-service stacks)
- Manage volumes or networks
- Debug a crashing container or analyze logs
- Check Docker disk usage or free up space
- Review or optimize a Dockerfile
## Prerequisites
- Docker Engine installed and running
- User added to the `docker` group (or use `sudo`)
- Docker Compose v2 (included with modern Docker installations)
Quick check:
```bash
docker --version && docker compose version
```
## Quick Reference
| Task | Command |
|------|---------|
| Run container (background) | `docker run -d --name NAME IMAGE` |
| Stop + remove | `docker stop NAME && docker rm NAME` |
| View logs (follow) | `docker logs --tail 50 -f NAME` |
| Shell into container | `docker exec -it NAME /bin/sh` |
| List all containers | `docker ps -a` |
| Build image | `docker build -t TAG .` |
| Compose up | `docker compose up -d` |
| Compose down | `docker compose down` |
| Disk usage | `docker system df` |
| Cleanup dangling | `docker image prune && docker container prune` |
## Procedure
### 1. Identify the domain
Figure out which area the request falls into:
- **Container lifecycle** → run, stop, start, restart, rm, pause/unpause
- **Container interaction** → exec, cp, logs, inspect, stats
- **Image management** → build, pull, push, tag, rmi, save/load
- **Docker Compose** → up, down, ps, logs, exec, build, config
- **Volumes & networks** → create, inspect, rm, prune, connect
- **Troubleshooting** → log analysis, exit codes, resource issues
### 2. Container operations
**Run a new container:**
```bash
# Detached service with port mapping
docker run -d --name web -p 8080:80 nginx
# With environment variables
docker run -d -e POSTGRES_PASSWORD=secret -e POSTGRES_DB=mydb --name db postgres:16
# With persistent data (named volume)
docker run -d -v pgdata:/var/lib/postgresql/data --name db postgres:16
# For development (bind mount source code)
docker run -d -v $(pwd)/src:/app/src -p 3000:3000 --name dev my-app
# Interactive debugging (auto-remove on exit)
docker run -it --rm ubuntu:22.04 /bin/bash
# With resource limits and restart policy
docker run -d --memory=512m --cpus=1.5 --restart=unless-stopped --name app my-app
```
Key flags: `-d` detached, `-it` interactive+tty, `--rm` auto-remove, `-p` port (host:container), `-e` env var, `-v` volume, `--name` name, `--restart` restart policy.
**Manage running containers:**
```bash
docker ps # running containers
docker ps -a # all (including stopped)
docker stop NAME # graceful stop
docker start NAME # start stopped container
docker restart NAME # stop + start
docker rm NAME # remove stopped container
docker rm -f NAME # force remove running container
docker container prune # remove ALL stopped containers
```
**Interact with containers:**
```bash
docker exec -it NAME /bin/sh # shell access (use /bin/bash if available)
docker exec NAME env # view environment variables
docker exec -u root NAME apt update # run as specific user
docker logs --tail 100 -f NAME # follow last 100 lines
docker logs --since 2h NAME # logs from last 2 hours
docker cp NAME:/path/file ./local # copy file from container
docker cp ./file NAME:/path/ # copy file to container
docker inspect NAME # full container details (JSON)
docker stats --no-stream # resource usage snapshot
docker top NAME # running processes
```
### 3. Image management
```bash
# Build
docker build -t my-app:latest .
docker build -t my-app:prod -f Dockerfile.prod .
docker build --no-cache -t my-app . # clean rebuild
DOCKER_BUILDKIT=1 docker build -t my-app . # faster with BuildKit
# Pull and push
docker pull node:20-alpine
docker login ghcr.io
docker tag my-app:latest registry/my-app:v1.0
docker push registry/my-app:v1.0
# Inspect
docker images # list local images
docker history IMAGE # see layers
docker inspect IMAGE # full details
# Cleanup
docker image prune # remove dangling (untagged) images
docker image prune -a # remove ALL unused images (careful!)
docker image prune -a --filter "until=168h" # unused images older than 7 days
```
### 4. Docker Compose
```bash
# Start/stop
docker compose up -d # start all services detached
docker compose up -d --build # rebuild images before starting
docker compose down # stop and remove containers
docker compose down -v # also remove volumes (DESTROYS DATA)
# Monitoring
docker compose ps # list services
docker compose logs -f api # follow logs for specific service
docker compose logs --tail 50 # last 50 lines all services
# Interaction
docker compose exec api /bin/sh # shell into running service
docker compose run --rm api npm test # one-off command (new container)
docker compose restart api # restart specific service
# Validation
docker compose config # validate and view resolved config
```
**Minimal compose.yml example:**
```yaml
services:
api:
build: .
ports:
- "3000:3000"
environment:
- DATABASE_URL=postgres://user:pass@db:5432/mydb
depends_on:
db:
condition: service_healthy
db:
image: postgres:16-alpine
environment:
POSTGRES_USER: user
POSTGRES_PASSWORD: pass
POSTGRES_DB: mydb
volumes:
- pgdata:/var/lib/postgresql/data
healthcheck:
test: ["CMD-SHELL", "pg_isready -U user"]
interval: 10s
timeout: 5s
retries: 5
volumes:
pgdata:
```
### 5. Volumes and networks
```bash
# Volumes
docker volume ls # list volumes
docker volume create mydata # create named volume
docker volume inspect mydata # details (mount point, etc.)
docker volume rm mydata # remove (fails if in use)
docker volume prune # remove unused volumes
# Networks
docker network ls # list networks
docker network create mynet # create bridge network
docker network inspect mynet # details (connected containers)
docker network connect mynet NAME # attach container to network
docker network disconnect mynet NAME # detach container
docker network rm mynet # remove network
docker network prune # remove unused networks
```
### 6. Disk usage and cleanup
Always start with a diagnostic before cleaning:
```bash
# Check what's using space
docker system df # summary
docker system df -v # detailed breakdown
# Targeted cleanup (safe)
docker container prune # stopped containers
docker image prune # dangling images
docker volume prune # unused volumes
docker network prune # unused networks
# Aggressive cleanup (confirm with user first!)
docker system prune # containers + images + networks
docker system prune -a # also unused images
docker system prune -a --volumes # EVERYTHING — named volumes too
```
**Warning:** Never run `docker system prune -a --volumes` without confirming with the user. This removes named volumes with potentially important data.
## Pitfalls
| Problem | Cause | Fix |
|---------|-------|-----|
| Container exits immediately | Main process finished or crashed | Check `docker logs NAME`, try `docker run -it --entrypoint /bin/sh IMAGE` |
| "port is already allocated" | Another process using that port | `docker ps` or `lsof -i :PORT` to find it |
| "no space left on device" | Docker disk full | `docker system df` then targeted prune |
| Can't connect to container | App binds to 127.0.0.1 inside container | App must bind to `0.0.0.0`, check `-p` mapping |
| Permission denied on volume | UID/GID mismatch host vs container | Use `--user $(id -u):$(id -g)` or fix permissions |
| Compose services can't reach each other | Wrong network or service name | Services use service name as hostname, check `docker compose config` |
| Build cache not working | Layer order wrong in Dockerfile | Put rarely-changing layers first (deps before source code) |
| Image too large | No multi-stage build, no .dockerignore | Use multi-stage builds, add `.dockerignore` |
## Verification
After any Docker operation, verify the result:
- **Container started?** → `docker ps` (check status is "Up")
- **Logs clean?** → `docker logs --tail 20 NAME` (no errors)
- **Port accessible?** → `curl -s http://localhost:PORT` or `docker port NAME`
- **Image built?** → `docker images | grep TAG`
- **Compose stack healthy?** → `docker compose ps` (all services "running" or "healthy")
- **Disk freed?** → `docker system df` (compare before/after)
## Dockerfile Optimization Tips
When reviewing or creating a Dockerfile, suggest these improvements:
1. **Multi-stage builds** — separate build environment from runtime to reduce final image size
2. **Layer ordering** — put dependencies before source code so changes don't invalidate cached layers
3. **Combine RUN commands** — fewer layers, smaller image
4. **Use .dockerignore** — exclude `node_modules`, `.git`, `__pycache__`, etc.
5. **Pin base image versions** — `node:20-alpine` not `node:latest`
6. **Run as non-root** — add `USER` instruction for security
7. **Use slim/alpine bases** — `python:3.12-slim` not `python:3.12`

208
website/docs/user-guide/skills/optional/dogfood/dogfood-adversarial-ux-test.md Normal file
View file

@ -0,0 +1,208 @@
---
title: "Adversarial Ux Test — Roleplay the most difficult, tech-resistant user for your product"
sidebar_label: "Adversarial Ux Test"
description: "Roleplay the most difficult, tech-resistant user for your product"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Adversarial Ux Test
Roleplay the most difficult, tech-resistant user for your product. Browse the app as that persona, find every UX pain point, then filter complaints through a pragmatism layer to separate real problems from noise. Creates actionable tickets from genuine issues only.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/dogfood/adversarial-ux-test` |
| Path | `optional-skills/dogfood/adversarial-ux-test` |
| Version | `1.0.0` |
| Author | Omni @ Comelse |
| License | MIT |
| Tags | `qa`, `ux`, `testing`, `adversarial`, `dogfood`, `personas`, `user-testing` |
| Related skills | [`dogfood`](/docs/user-guide/skills/bundled/dogfood/dogfood-dogfood) |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Adversarial UX Test
Roleplay the worst-case user for your product — the person who hates technology, doesn't want your software, and will find every reason to complain. Then filter their feedback through a pragmatism layer to separate real UX problems from "I hate computers" noise.
Think of it as an automated "mom test" — but angry.
## Why This Works
Most QA finds bugs. This finds **friction**. A technically correct app can still be unusable for real humans. The adversarial persona catches:
- Confusing terminology that makes sense to developers but not users
- Too many steps to accomplish basic tasks
- Missing onboarding or "aha moments"
- Accessibility issues (font size, contrast, click targets)
- Cold-start problems (empty states, no demo content)
- Paywall/signup friction that kills conversion
The **pragmatism filter** (Phase 3) is what makes this useful instead of just entertaining. Without it, you'd add a "print this page" button to every screen because Grandpa can't figure out PDFs.
## How to Use
Tell the agent:
```
"Run an adversarial UX test on [URL]"
"Be a grumpy [persona type] and test [app name]"
"Do an asshole user test on my staging site"
```
You can provide a persona or let the agent generate one based on your product's target audience.
## Step 1: Define the Persona
If no persona is provided, generate one by answering:
1. **Who is the HARDEST user for this product?** (age 50+, non-technical role, decades of experience doing it "the old way")
2. **What is their tech comfort level?** (the lower the better — WhatsApp-only, paper notebooks, wife set up their email)
3. **What is the ONE thing they need to accomplish?** (their core job, not your feature list)
4. **What would make them give up?** (too many clicks, jargon, slow, confusing)
5. **How do they talk when frustrated?** (blunt, sweary, dismissive, sighing)
### Good Persona Example
> **"Big Mick" McAllister** — 58-year-old S&C coach. Uses WhatsApp and that's it. His "spreadsheet" is a paper notebook. "If I can't figure it out in 10 seconds I'm going back to my notebook." Needs to log session results for 25 players. Hates small text, jargon, and passwords.
### Bad Persona Example
> "A user who doesn't like the app" — too vague, no constraints, no voice.
The persona must be **specific enough to stay in character** for 20 minutes of testing.
## Step 2: Become the Asshole (Browse as the Persona)
1. Read any available project docs for app context and URLs
2. **Fully inhabit the persona** — their frustrations, limitations, goals
3. Navigate to the app using browser tools
4. **Attempt the persona's ACTUAL TASKS** (not a feature tour):
- Can they do what they came to do?
- How many clicks/screens to accomplish it?
- What confuses them?
- What makes them angry?
- Where do they get lost?
- What would make them give up and go back to their old way?
5. Test these friction categories:
- **First impression** — would they even bother past the landing page?
- **Core workflow** — the ONE thing they need to do most often
- **Error recovery** — what happens when they do something wrong?
- **Readability** — text size, contrast, information density
- **Speed** — does it feel faster than their current method?
- **Terminology** — any jargon they wouldn't understand?
- **Navigation** — can they find their way back? do they know where they are?
6. Take screenshots of every pain point
7. Check browser console for JS errors on every page
## Step 3: The Rant (Write Feedback in Character)
Write the feedback AS THE PERSONA — in their voice, with their frustrations. This is not a bug report. This is a real human venting.
```
[PERSONA NAME]'s Review of [PRODUCT]
Overall: [Would they keep using it? Yes/No/Maybe with conditions]
THE GOOD (grudging admission):
- [things even they have to admit work]
THE BAD (legitimate UX issues):
- [real problems that would stop them from using the product]
THE UGLY (showstoppers):
- [things that would make them uninstall/cancel immediately]
SPECIFIC COMPLAINTS:
1. [Page/feature]: "[quote in persona voice]" — [what happened, expected]
2. ...
VERDICT: "[one-line persona quote summarizing their experience]"
```
## Step 4: The Pragmatism Filter (Critical — Do Not Skip)
Step OUT of the persona. Evaluate each complaint as a product person:
- **RED: REAL UX BUG** — Any user would have this problem, not just grumpy ones. Fix it.
- **YELLOW: VALID BUT LOW PRIORITY** — Real issue but only for extreme users. Note it.
- **WHITE: PERSONA NOISE** — "I hate computers" talking, not a product problem. Skip it.
- **GREEN: FEATURE REQUEST** — Good idea hidden in the complaint. Consider it.
### Filter Criteria
1. Would a 35-year-old competent-but-busy user have the same complaint? → RED
2. Is this a genuine accessibility issue (font size, contrast, click targets)? → RED
3. Is this "I want it to work like paper" resistance to digital? → WHITE
4. Is this a real workflow inefficiency the persona stumbled on? → YELLOW or RED
5. Would fixing this add complexity for the 80% who are fine? → WHITE
6. Does the complaint reveal a missing onboarding moment? → GREEN
**This filter is MANDATORY.** Never ship raw persona complaints as tickets.
## Step 5: Create Tickets
For **RED** and **GREEN** items only:
- Clear, actionable title
- Include the persona's verbatim quote (entertaining + memorable)
- The real UX issue underneath (objective)
- A suggested fix (actionable)
- Tag/label: "ux-review"
For **YELLOW** items: one catch-all ticket with all notes.
**WHITE** items appear in the report only. No tickets.
**Max 10 tickets per session** — focus on the worst issues.
## Step 6: Report
Deliver:
1. The persona rant (Step 3) — entertaining and visceral
2. The filtered assessment (Step 4) — pragmatic and actionable
3. Tickets created (Step 5) — with links
4. Screenshots of key issues
## Tips
- **One persona per session.** Don't mix perspectives.
- **Stay in character during Steps 2-3.** Break character only at Step 4.
- **Test the CORE WORKFLOW first.** Don't get distracted by settings pages.
- **Empty states are gold.** New user experience reveals the most friction.
- **The best findings are RED items the persona found accidentally** while trying to do something else.
- **If the persona has zero complaints, your persona is too tech-savvy.** Make them older, less patient, more set in their ways.
- **Run this before demos, launches, or after shipping a batch of features.**
- **Register as a NEW user when possible.** Don't use pre-seeded admin accounts — the cold start experience is where most friction lives.
- **Zero WHITE items is a signal, not a failure.** If the pragmatism filter finds no noise, your product has real UX problems, not just a grumpy persona.
- **Check known issues in project docs AFTER the test.** If the persona found a bug that's already in the known issues list, that's actually the most damning finding — it means the team knew about it but never felt the user's pain.
- **Subscription/paywall testing is critical.** Test with expired accounts, not just active ones. The "what happens when you can't pay" experience reveals whether the product respects users or holds their data hostage.
- **Count the clicks to accomplish the persona's ONE task.** If it's more than 5, that's almost always a RED finding regardless of persona tech level.
## Example Personas by Industry
These are starting points — customize for your specific product:
| Product Type | Persona | Age | Key Trait |
|-------------|---------|-----|-----------|
| CRM | Retirement home director | 68 | Filing cabinet is the current CRM |
| Photography SaaS | Rural wedding photographer | 62 | Books clients by phone, invoices on paper |
| AI/ML Tool | Department store buyer | 55 | Burned by 3 failed tech startups |
| Fitness App | Old-school gym coach | 58 | Paper notebook, thick fingers, bad eyes |
| Accounting | Family bakery owner | 64 | Shoebox of receipts, hates subscriptions |
| E-commerce | Market stall vendor | 60 | Cash only, smartphone is for calls |
| Healthcare | Senior GP | 63 | Dictates notes, nurse handles the computer |
| Education | Veteran teacher | 57 | Chalk and talk, worksheets in ring binders |
## Rules
- Stay in character during Steps 2-3
- Be genuinely mean but fair — find real problems, not manufactured ones
- The pragmatism filter (Step 4) is **MANDATORY**
- Screenshots required for every complaint
- Max 10 tickets per session
- Test on staging/deployed app, not local dev
- One persona, one session, one report

142
website/docs/user-guide/skills/optional/email/email-agentmail.md Normal file
View file

@ -0,0 +1,142 @@
---
title: "Agentmail — Give the agent its own dedicated email inbox via AgentMail"
sidebar_label: "Agentmail"
description: "Give the agent its own dedicated email inbox via AgentMail"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Agentmail
Give the agent its own dedicated email inbox via AgentMail. Send, receive, and manage email autonomously using agent-owned email addresses (e.g. hermes-agent@agentmail.to).
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/email/agentmail` |
| Path | `optional-skills/email/agentmail` |
| Version | `1.0.0` |
| Tags | `email`, `communication`, `agentmail`, `mcp` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# AgentMail — Agent-Owned Email Inboxes
## Requirements
- **AgentMail API key** (required) — sign up at https://console.agentmail.to (free tier: 3 inboxes, 3,000 emails/month; paid plans from $20/mo)
- Node.js 18+ (for the MCP server)
## When to Use
Use this skill when you need to:
- Give the agent its own dedicated email address
- Send emails autonomously on behalf of the agent
- Receive and read incoming emails
- Manage email threads and conversations
- Sign up for services or authenticate via email
- Communicate with other agents or humans via email
This is NOT for reading the user's personal email (use himalaya or Gmail for that).
AgentMail gives the agent its own identity and inbox.
## Setup
### 1. Get an API Key
- Go to https://console.agentmail.to
- Create an account and generate an API key (starts with `am_`)
### 2. Configure MCP Server
Add to `~/.hermes/config.yaml` (paste your actual key — MCP env vars are not expanded from .env):
```yaml
mcp_servers:
agentmail:
command: "npx"
args: ["-y", "agentmail-mcp"]
env:
AGENTMAIL_API_KEY: "am_your_key_here"
```
### 3. Restart Hermes
```bash
hermes
```
All 11 AgentMail tools are now available automatically.
## Available Tools (via MCP)
| Tool | Description |
|------|-------------|
| `list_inboxes` | List all agent inboxes |
| `get_inbox` | Get details of a specific inbox |
| `create_inbox` | Create a new inbox (gets a real email address) |
| `delete_inbox` | Delete an inbox |
| `list_threads` | List email threads in an inbox |
| `get_thread` | Get a specific email thread |
| `send_message` | Send a new email |
| `reply_to_message` | Reply to an existing email |
| `forward_message` | Forward an email |
| `update_message` | Update message labels/status |
| `get_attachment` | Download an email attachment |
## Procedure
### Create an inbox and send an email
1. Create a dedicated inbox:
- Use `create_inbox` with a username (e.g. `hermes-agent`)
- The agent gets address: `hermes-agent@agentmail.to`
2. Send an email:
- Use `send_message` with `inbox_id`, `to`, `subject`, `text`
3. Check for replies:
- Use `list_threads` to see incoming conversations
- Use `get_thread` to read a specific thread
### Check incoming email
1. Use `list_inboxes` to find your inbox ID
2. Use `list_threads` with the inbox ID to see conversations
3. Use `get_thread` to read a thread and its messages
### Reply to an email
1. Get the thread with `get_thread`
2. Use `reply_to_message` with the message ID and your reply text
## Example Workflows
**Sign up for a service:**
```
1. create_inbox (username: "signup-bot")
2. Use the inbox address to register on the service
3. list_threads to check for verification email
4. get_thread to read the verification code
```
**Agent-to-human outreach:**
```
1. create_inbox (username: "hermes-outreach")
2. send_message (to: user@example.com, subject: "Hello", text: "...")
3. list_threads to check for replies
```
## Pitfalls
- Free tier limited to 3 inboxes and 3,000 emails/month
- Emails come from `@agentmail.to` domain on free tier (custom domains on paid plans)
- Node.js (18+) is required for the MCP server (`npx -y agentmail-mcp`)
- The `mcp` Python package must be installed: `pip install mcp`
- Real-time inbound email (webhooks) requires a public server — use `list_threads` polling via cronjob instead for personal use
## Verification
After setup, test with:
```
hermes --toolsets mcp -q "Create an AgentMail inbox called test-agent and tell me its email address"
```
You should see the new inbox address returned.
## References
- AgentMail docs: https://docs.agentmail.to/
- AgentMail console: https://console.agentmail.to
- AgentMail MCP repo: https://github.com/agentmail-to/agentmail-mcp
- Pricing: https://www.agentmail.to/pricing

257
website/docs/user-guide/skills/optional/health/health-fitness-nutrition.md Normal file
View file

@ -0,0 +1,257 @@
---
title: "Fitness Nutrition — Gym workout planner and nutrition tracker"
sidebar_label: "Fitness Nutrition"
description: "Gym workout planner and nutrition tracker"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Fitness Nutrition
Gym workout planner and nutrition tracker. Search 690+ exercises by muscle, equipment, or category via wger. Look up macros and calories for 380,000+ foods via USDA FoodData Central. Compute BMI, TDEE, one-rep max, macro splits, and body fat — pure Python, no pip installs. Built for anyone chasing gains, cutting weight, or just trying to eat better.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/health/fitness-nutrition` |
| Path | `optional-skills/health/fitness-nutrition` |
| Version | `1.0.0` |
| License | MIT |
| Tags | `health`, `fitness`, `nutrition`, `gym`, `workout`, `diet`, `exercise` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Fitness & Nutrition
Expert fitness coach and sports nutritionist skill. Two data sources
plus offline calculators — everything a gym-goer needs in one place.
**Data sources (all free, no pip dependencies):**
- **wger** (https://wger.de/api/v2/) — open exercise database, 690+ exercises with muscles, equipment, images. Public endpoints need zero authentication.
- **USDA FoodData Central** (https://api.nal.usda.gov/fdc/v1/) — US government nutrition database, 380,000+ foods. `DEMO_KEY` works instantly; free signup for higher limits.
**Offline calculators (pure stdlib Python):**
- BMI, TDEE (Mifflin-St Jeor), one-rep max (Epley/Brzycki/Lombardi), macro splits, body fat % (US Navy method)
---
## When to Use
Trigger this skill when the user asks about:
- Exercises, workouts, gym routines, muscle groups, workout splits
- Food macros, calories, protein content, meal planning, calorie counting
- Body composition: BMI, body fat, TDEE, caloric surplus/deficit
- One-rep max estimates, training percentages, progressive overload
- Macro ratios for cutting, bulking, or maintenance
---
## Procedure
### Exercise Lookup (wger API)
All wger public endpoints return JSON and require no auth. Always add
`format=json` and `language=2` (English) to exercise queries.
**Step 1 — Identify what the user wants:**
- By muscle → use `/api/v2/exercise/?muscles={id}&language=2&status=2&format=json`
- By category → use `/api/v2/exercise/?category={id}&language=2&status=2&format=json`
- By equipment → use `/api/v2/exercise/?equipment={id}&language=2&status=2&format=json`
- By name → use `/api/v2/exercise/search/?term={query}&language=english&format=json`
- Full details → use `/api/v2/exerciseinfo/{exercise_id}/?format=json`
**Step 2 — Reference IDs (so you don't need extra API calls):**
Exercise categories:
| ID | Category |
|----|-------------|
| 8 | Arms |
| 9 | Legs |
| 10 | Abs |
| 11 | Chest |
| 12 | Back |
| 13 | Shoulders |
| 14 | Calves |
| 15 | Cardio |
Muscles:
| ID | Muscle | ID | Muscle |
|----|---------------------------|----|-------------------------|
| 1 | Biceps brachii | 2 | Anterior deltoid |
| 3 | Serratus anterior | 4 | Pectoralis major |
| 5 | Obliquus externus | 6 | Gastrocnemius |
| 7 | Rectus abdominis | 8 | Gluteus maximus |
| 9 | Trapezius | 10 | Quadriceps femoris |
| 11 | Biceps femoris | 12 | Latissimus dorsi |
| 13 | Brachialis | 14 | Triceps brachii |
| 15 | Soleus | | |
Equipment:
| ID | Equipment |
|----|----------------|
| 1 | Barbell |
| 3 | Dumbbell |
| 4 | Gym mat |
| 5 | Swiss Ball |
| 6 | Pull-up bar |
| 7 | none (bodyweight) |
| 8 | Bench |
| 9 | Incline bench |
| 10 | Kettlebell |
**Step 3 — Fetch and present results:**
```bash
# Search exercises by name
QUERY="$1"
ENCODED=$(python3 -c "import urllib.parse,sys; print(urllib.parse.quote(sys.argv[1]))" "$QUERY")
curl -s "https://wger.de/api/v2/exercise/search/?term=${ENCODED}&language=english&format=json" \
| python3 -c "
import json,sys
data=json.load(sys.stdin)
for s in data.get('suggestions',[])[:10]:
d=s.get('data',{})
print(f\" ID {d.get('id','?'):>4} | {d.get('name','N/A'):<35} | Category: {d.get('category','N/A')}\")
"
```
```bash
# Get full details for a specific exercise
EXERCISE_ID="$1"
curl -s "https://wger.de/api/v2/exerciseinfo/${EXERCISE_ID}/?format=json" \
| python3 -c "
import json,sys,html,re
data=json.load(sys.stdin)
trans=[t for t in data.get('translations',[]) if t.get('language')==2]
t=trans[0] if trans else data.get('translations',[{}])[0]
desc=re.sub('<[^>]+>','',html.unescape(t.get('description','N/A')))
print(f\"Exercise : {t.get('name','N/A')}\")
print(f\"Category : {data.get('category',{}).get('name','N/A')}\")
print(f\"Primary : {', '.join(m.get('name_en','') for m in data.get('muscles',[])) or 'N/A'}\")
print(f\"Secondary : {', '.join(m.get('name_en','') for m in data.get('muscles_secondary',[])) or 'none'}\")
print(f\"Equipment : {', '.join(e.get('name','') for e in data.get('equipment',[])) or 'bodyweight'}\")
print(f\"How to : {desc[:500]}\")
imgs=data.get('images',[])
if imgs: print(f\"Image : {imgs[0].get('image','')}\")
"
```
```bash
# List exercises filtering by muscle, category, or equipment
# Combine filters as needed: ?muscles=4&equipment=1&language=2&status=2
FILTER="$1" # e.g. "muscles=4" or "category=11" or "equipment=3"
curl -s "https://wger.de/api/v2/exercise/?${FILTER}&language=2&status=2&limit=20&format=json" \
| python3 -c "
import json,sys
data=json.load(sys.stdin)
print(f'Found {data.get(\"count\",0)} exercises.')
for ex in data.get('results',[]):
print(f\" ID {ex['id']:>4} | muscles: {ex.get('muscles',[])} | equipment: {ex.get('equipment',[])}\")
"
```
### Nutrition Lookup (USDA FoodData Central)
Uses `USDA_API_KEY` env var if set, otherwise falls back to `DEMO_KEY`.
DEMO_KEY = 30 requests/hour. Free signup key = 1,000 requests/hour.
```bash
# Search foods by name
FOOD="$1"
API_KEY="${USDA_API_KEY:-DEMO_KEY}"
ENCODED=$(python3 -c "import urllib.parse,sys; print(urllib.parse.quote(sys.argv[1]))" "$FOOD")
curl -s "https://api.nal.usda.gov/fdc/v1/foods/search?api_key=${API_KEY}&query=${ENCODED}&pageSize=5&dataType=Foundation,SR%20Legacy" \
| python3 -c "
import json,sys
data=json.load(sys.stdin)
foods=data.get('foods',[])
if not foods: print('No foods found.'); sys.exit()
for f in foods:
n={x['nutrientName']:x.get('value','?') for x in f.get('foodNutrients',[])}
cal=n.get('Energy','?'); prot=n.get('Protein','?')
fat=n.get('Total lipid (fat)','?'); carb=n.get('Carbohydrate, by difference','?')
print(f\"{f.get('description','N/A')}\")
print(f\" Per 100g: {cal} kcal | {prot}g protein | {fat}g fat | {carb}g carbs\")
print(f\" FDC ID: {f.get('fdcId','N/A')}\")
print()
"
```
```bash
# Detailed nutrient profile by FDC ID
FDC_ID="$1"
API_KEY="${USDA_API_KEY:-DEMO_KEY}"
curl -s "https://api.nal.usda.gov/fdc/v1/food/${FDC_ID}?api_key=${API_KEY}" \
| python3 -c "
import json,sys
d=json.load(sys.stdin)
print(f\"Food: {d.get('description','N/A')}\")
print(f\"{'Nutrient':<40} {'Amount':>8} {'Unit'}\")
print('-'*56)
for x in sorted(d.get('foodNutrients',[]),key=lambda x:x.get('nutrient',{}).get('rank',9999)):
nut=x.get('nutrient',{}); amt=x.get('amount',0)
if amt and float(amt)>0:
print(f\" {nut.get('name',''):<38} {amt:>8} {nut.get('unitName','')}\")
"
```
### Offline Calculators
Use the helper scripts in `scripts/` for batch operations,
or run inline for single calculations:
- `python3 scripts/body_calc.py bmi <weight_kg> <height_cm>`
- `python3 scripts/body_calc.py tdee <weight_kg> <height_cm> <age> <M|F> <activity 1-5>`
- `python3 scripts/body_calc.py 1rm <weight> <reps>`
- `python3 scripts/body_calc.py macros <tdee_kcal> <cut|maintain|bulk>`
- `python3 scripts/body_calc.py bodyfat <M|F> <neck_cm> <waist_cm> [hip_cm] <height_cm>`
See `references/FORMULAS.md` for the science behind each formula.
---
## Pitfalls
- wger exercise endpoint returns **all languages by default** — always add `language=2` for English
- wger includes **unverified user submissions** — add `status=2` to only get approved exercises
- USDA `DEMO_KEY` has **30 req/hour** — add `sleep 2` between batch requests or get a free key
- USDA data is **per 100g** — remind users to scale to their actual portion size
- BMI does not distinguish muscle from fat — high BMI in muscular people is not necessarily unhealthy
- Body fat formulas are **estimates** (±3-5%) — recommend DEXA scans for precision
- 1RM formulas lose accuracy above 10 reps — use sets of 3-5 for best estimates
- wger's `exercise/search` endpoint uses `term` not `query` as the parameter name
---
## Verification
After running exercise search: confirm results include exercise names, muscle groups, and equipment.
After nutrition lookup: confirm per-100g macros are returned with kcal, protein, fat, carbs.
After calculators: sanity-check outputs (e.g. TDEE should be 1500-3500 for most adults).
---
## Quick Reference
| Task | Source | Endpoint |
|------|--------|----------|
| Search exercises by name | wger | `GET /api/v2/exercise/search/?term=&language=english` |
| Exercise details | wger | `GET /api/v2/exerciseinfo/{id}/` |
| Filter by muscle | wger | `GET /api/v2/exercise/?muscles={id}&language=2&status=2` |
| Filter by equipment | wger | `GET /api/v2/exercise/?equipment={id}&language=2&status=2` |
| List categories | wger | `GET /api/v2/exercisecategory/` |
| List muscles | wger | `GET /api/v2/muscle/` |
| Search foods | USDA | `GET /fdc/v1/foods/search?query=&dataType=Foundation,SR Legacy` |
| Food details | USDA | `GET /fdc/v1/food/{fdcId}` |
| BMI / TDEE / 1RM / macros | offline | `python3 scripts/body_calc.py` |

469
website/docs/user-guide/skills/optional/health/health-neuroskill-bci.md Normal file
View file

@ -0,0 +1,469 @@
---
title: "Neuroskill Bci"
sidebar_label: "Neuroskill Bci"
description: "Connect to a running NeuroSkill instance and incorporate the user's real-time cognitive and emotional state (focus, relaxation, mood, cognitive load, drowsin..."
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Neuroskill Bci
Connect to a running NeuroSkill instance and incorporate the user's real-time cognitive and emotional state (focus, relaxation, mood, cognitive load, drowsiness, heart rate, HRV, sleep staging, and 40+ derived EXG scores) into responses. Requires a BCI wearable (Muse 2/S or OpenBCI) and the NeuroSkill desktop app running locally.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/health/neuroskill-bci` |
| Path | `optional-skills/health/neuroskill-bci` |
| Version | `1.0.0` |
| Author | Hermes Agent + Nous Research |
| License | MIT |
| Tags | `BCI`, `neurofeedback`, `health`, `focus`, `EEG`, `cognitive-state`, `biometrics`, `neuroskill` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# NeuroSkill BCI Integration
Connect Hermes to a running [NeuroSkill](https://neuroskill.com/) instance to read
real-time brain and body metrics from a BCI wearable. Use this to give
cognitively-aware responses, suggest interventions, and track mental performance
over time.
> **⚠️ Research Use Only** — NeuroSkill is an open-source research tool. It is
> NOT a medical device and has NOT been cleared by the FDA, CE, or any regulatory
> body. Never use these metrics for clinical diagnosis or treatment.
See `references/metrics.md` for the full metric reference, `references/protocols.md`
for intervention protocols, and `references/api.md` for the WebSocket/HTTP API.
---
## Prerequisites
- **Node.js 20+** installed (`node --version`)
- **NeuroSkill desktop app** running with a connected BCI device
- **BCI hardware**: Muse 2, Muse S, or OpenBCI (4-channel EEG + PPG + IMU via BLE)
- `npx neuroskill status` returns data without errors
### Verify Setup
```bash
node --version # Must be 20+
npx neuroskill status # Full system snapshot
npx neuroskill status --json # Machine-parseable JSON
```
If `npx neuroskill status` returns an error, tell the user:
- Make sure the NeuroSkill desktop app is open
- Ensure the BCI device is powered on and connected via Bluetooth
- Check signal quality — green indicators in NeuroSkill (≥0.7 per electrode)
- If `command not found`, install Node.js 20+
---
## CLI Reference: `npx neuroskill <command>`
All commands support `--json` (raw JSON, pipe-safe) and `--full` (human summary + JSON).
| Command | Description |
|---------|-------------|
| `status` | Full system snapshot: device, scores, bands, ratios, sleep, history |
| `session [N]` | Single session breakdown with first/second half trends (0=most recent) |
| `sessions` | List all recorded sessions across all days |
| `search` | ANN similarity search for neurally similar historical moments |
| `compare` | A/B session comparison with metric deltas and trend analysis |
| `sleep [N]` | Sleep stage classification (Wake/N1/N2/N3/REM) with analysis |
| `label "text"` | Create a timestamped annotation at the current moment |
| `search-labels "query"` | Semantic vector search over past labels |
| `interactive "query"` | Cross-modal 4-layer graph search (text → EXG → labels) |
| `listen` | Real-time event streaming (default 5s, set `--seconds N`) |
| `umap` | 3D UMAP projection of session embeddings |
| `calibrate` | Open calibration window and start a profile |
| `timer` | Launch focus timer (Pomodoro/Deep Work/Short Focus presets) |
| `notify "title" "body"` | Send an OS notification via the NeuroSkill app |
| `raw '{json}'` | Raw JSON passthrough to the server |
### Global Flags
| Flag | Description |
|------|-------------|
| `--json` | Raw JSON output (no ANSI, pipe-safe) |
| `--full` | Human summary + colorized JSON |
| `--port <N>` | Override server port (default: auto-discover, usually 8375) |
| `--ws` | Force WebSocket transport |
| `--http` | Force HTTP transport |
| `--k <N>` | Nearest neighbors count (search, search-labels) |
| `--seconds <N>` | Duration for listen (default: 5) |
| `--trends` | Show per-session metric trends (sessions) |
| `--dot` | Graphviz DOT output (interactive) |
---
## 1. Checking Current State
### Get Live Metrics
```bash
npx neuroskill status --json
```
**Always use `--json`** for reliable parsing. The default output is colorized
human-readable text.
### Key Fields in the Response
The `scores` object contains all live metrics (0–1 scale unless noted):
```jsonc
{
"scores": {
"focus": 0.70, // β / (α + θ) — sustained attention
"relaxation": 0.40, // α / (β + θ) — calm wakefulness
"engagement": 0.60, // active mental investment
"meditation": 0.52, // alpha + stillness + HRV coherence
"mood": 0.55, // composite from FAA, TAR, BAR
"cognitive_load": 0.33, // frontal θ / temporal α · f(FAA, TBR)
"drowsiness": 0.10, // TAR + TBR + falling spectral centroid
"hr": 68.2, // heart rate in bpm (from PPG)
"snr": 14.3, // signal-to-noise ratio in dB
"stillness": 0.88, // 0–1; 1 = perfectly still
"faa": 0.042, // Frontal Alpha Asymmetry (+ = approach)
"tar": 0.56, // Theta/Alpha Ratio
"bar": 0.53, // Beta/Alpha Ratio
"tbr": 1.06, // Theta/Beta Ratio (ADHD proxy)
"apf": 10.1, // Alpha Peak Frequency in Hz
"coherence": 0.614, // inter-hemispheric coherence
"bands": {
"rel_delta": 0.28, "rel_theta": 0.18,
"rel_alpha": 0.32, "rel_beta": 0.17, "rel_gamma": 0.05
}
}
}
```
Also includes: `device` (state, battery, firmware), `signal_quality` (per-electrode 0–1),
`session` (duration, epochs), `embeddings`, `labels`, `sleep` summary, and `history`.
### Interpreting the Output
Parse the JSON and translate metrics into natural language. Never report raw
numbers alone — always give them meaning:
**DO:**
> "Your focus is solid right now at 0.70 — that's flow state territory. Heart
> rate is steady at 68 bpm and your FAA is positive, which suggests good
> approach motivation. Great time to tackle something complex."
**DON'T:**
> "Focus: 0.70, Relaxation: 0.40, HR: 68"
Key interpretation thresholds (see `references/metrics.md` for the full guide):
- **Focus > 0.70** → flow state territory, protect it
- **Focus &lt; 0.40** → suggest a break or protocol
- **Drowsiness > 0.60** → fatigue warning, micro-sleep risk
- **Relaxation &lt; 0.30** → stress intervention needed
- **Cognitive Load > 0.70 sustained** → mind dump or break
- **TBR > 1.5** → theta-dominant, reduced executive control
- **FAA &lt; 0** → withdrawal/negative affect — consider FAA rebalancing
- **SNR &lt; 3 dB** → unreliable signal, suggest electrode repositioning
---
## 2. Session Analysis
### Single Session Breakdown
```bash
npx neuroskill session --json # most recent session
npx neuroskill session 1 --json # previous session
npx neuroskill session 0 --json | jq '{focus: .metrics.focus, trend: .trends.focus}'
```
Returns full metrics with **first-half vs second-half trends** (`"up"`, `"down"`, `"flat"`).
Use this to describe how a session evolved:
> "Your focus started at 0.64 and climbed to 0.76 by the end — a clear upward trend.
> Cognitive load dropped from 0.38 to 0.28, suggesting the task became more automatic
> as you settled in."
### List All Sessions
```bash
npx neuroskill sessions --json
npx neuroskill sessions --trends # show per-session metric trends
```
---
## 3. Historical Search
### Neural Similarity Search
```bash
npx neuroskill search --json # auto: last session, k=5
npx neuroskill search --k 10 --json # 10 nearest neighbors
npx neuroskill search --start <UTC> --end <UTC> --json
```
Finds moments in history that are neurally similar using HNSW approximate
nearest-neighbor search over 128-D ZUNA embeddings. Returns distance statistics,
temporal distribution (hour of day), and top matching days.
Use this when the user asks:
- "When was I last in a state like this?"
- "Find my best focus sessions"
- "When do I usually crash in the afternoon?"
### Semantic Label Search
```bash
npx neuroskill search-labels "deep focus" --k 10 --json
npx neuroskill search-labels "stress" --json | jq '[.results[].EXG_metrics.tbr]'
```
Searches label text using vector embeddings (Xenova/bge-small-en-v1.5). Returns
matching labels with their associated EXG metrics at the time of labeling.
### Cross-Modal Graph Search
```bash
npx neuroskill interactive "deep focus" --json
npx neuroskill interactive "deep focus" --dot | dot -Tsvg > graph.svg
```
4-layer graph: query → text labels → EXG points → nearby labels. Use `--k-text`,
`--k-EXG`, `--reach <minutes>` to tune.
---
## 4. Session Comparison
```bash
npx neuroskill compare --json # auto: last 2 sessions
npx neuroskill compare --a-start <UTC> --a-end <UTC> --b-start <UTC> --b-end <UTC> --json
```
Returns metric deltas with absolute change, percentage change, and direction for
~50 metrics. Also includes `insights.improved[]` and `insights.declined[]` arrays,
sleep staging for both sessions, and a UMAP job ID.
Interpret comparisons with context — mention trends, not just deltas:
> "Yesterday you had two strong focus blocks (10am and 2pm). Today you've had one
> starting around 11am that's still going. Your overall engagement is higher today
> but there have been more stress spikes — your stress index jumped 15% and
> FAA dipped negative more often."
```bash
# Sort metrics by improvement percentage
npx neuroskill compare --json | jq '.insights.deltas | to_entries | sort_by(.value.pct) | reverse'
```
---
## 5. Sleep Data
```bash
npx neuroskill sleep --json # last 24 hours
npx neuroskill sleep 0 --json # most recent sleep session
npx neuroskill sleep --start <UTC> --end <UTC> --json
```
Returns epoch-by-epoch sleep staging (5-second windows) with analysis:
- **Stage codes**: 0=Wake, 1=N1, 2=N2, 3=N3 (deep), 4=REM
- **Analysis**: efficiency_pct, onset_latency_min, rem_latency_min, bout counts
- **Healthy targets**: N3 15–25%, REM 20–25%, efficiency >85%, onset &lt;20 min
```bash
npx neuroskill sleep --json | jq '.summary | {n3: .n3_epochs, rem: .rem_epochs}'
npx neuroskill sleep --json | jq '.analysis.efficiency_pct'
```
Use this when the user mentions sleep, tiredness, or recovery.
---
## 6. Labeling Moments
```bash
npx neuroskill label "breakthrough"
npx neuroskill label "studying algorithms"
npx neuroskill label "post-meditation"
npx neuroskill label --json "focus block start" # returns label_id
```
Auto-label moments when:
- User reports a breakthrough or insight
- User starts a new task type (e.g., "switching to code review")
- User completes a significant protocol
- User asks you to mark the current moment
- A notable state transition occurs (entering/leaving flow)
Labels are stored in a database and indexed for later retrieval via `search-labels`
and `interactive` commands.
---
## 7. Real-Time Streaming
```bash
npx neuroskill listen --seconds 30 --json
npx neuroskill listen --seconds 5 --json | jq '[.[] | select(.event == "scores")]'
```
Streams live WebSocket events (EXG, PPG, IMU, scores, labels) for the specified
duration. Requires WebSocket connection (not available with `--http`).
Use this for continuous monitoring scenarios or to observe metric changes in real-time
during a protocol.
---
## 8. UMAP Visualization
```bash
npx neuroskill umap --json # auto: last 2 sessions
npx neuroskill umap --a-start <UTC> --a-end <UTC> --b-start <UTC> --b-end <UTC> --json
```
GPU-accelerated 3D UMAP projection of ZUNA embeddings. The `separation_score`
indicates how neurally distinct two sessions are:
- **> 1.5** → Sessions are neurally distinct (different brain states)
- **&lt; 0.5** → Similar brain states across both sessions
---
## 9. Proactive State Awareness
### Session Start Check
At the beginning of a session, optionally run a status check if the user mentions
they're wearing their device or asks about their state:
```bash
npx neuroskill status --json
```
Inject a brief state summary:
> "Quick check-in: focus is building at 0.62, relaxation is good at 0.55, and your
> FAA is positive — approach motivation is engaged. Looks like a solid start."
### When to Proactively Mention State
Mention cognitive state **only** when:
- User explicitly asks ("How am I doing?", "Check my focus")
- User reports difficulty concentrating, stress, or fatigue
- A critical threshold is crossed (drowsiness > 0.70, focus &lt; 0.30 sustained)
- User is about to do something cognitively demanding and asks for readiness
**Do NOT** interrupt flow state to report metrics. If focus > 0.75, protect the
session — silence is the correct response.
---
## 10. Suggesting Protocols
When metrics indicate a need, suggest a protocol from `references/protocols.md`.
Always ask before starting — never interrupt flow state:
> "Your focus has been declining for the past 15 minutes and TBR is climbing past
> 1.5 — signs of theta dominance and mental fatigue. Want me to walk you through
> a Theta-Beta Neurofeedback Anchor? It's a 90-second exercise that uses rhythmic
> counting and breath to suppress theta and lift beta."
Key triggers:
- **Focus &lt; 0.40, TBR > 1.5** → Theta-Beta Neurofeedback Anchor or Box Breathing
- **Relaxation &lt; 0.30, stress_index high** → Cardiac Coherence or 4-7-8 Breathing
- **Cognitive Load > 0.70 sustained** → Cognitive Load Offload (mind dump)
- **Drowsiness > 0.60** → Ultradian Reset or Wake Reset
- **FAA &lt; 0 (negative)** → FAA Rebalancing
- **Flow State (focus > 0.75, engagement > 0.70)** → Do NOT interrupt
- **High stillness + headache_index** → Neck Release Sequence
- **Low RMSSD (&lt; 25ms)** → Vagal Toning
---
## 11. Additional Tools
### Focus Timer
```bash
npx neuroskill timer --json
```
Launches the Focus Timer window with Pomodoro (25/5), Deep Work (50/10), or
Short Focus (15/5) presets.
### Calibration
```bash
npx neuroskill calibrate
npx neuroskill calibrate --profile "Eyes Open"
```
Opens the calibration window. Useful when signal quality is poor or the user
wants to establish a personalized baseline.
### OS Notifications
```bash
npx neuroskill notify "Break Time" "Your focus has been declining for 20 minutes"
```
### Raw JSON Passthrough
```bash
npx neuroskill raw '{"command":"status"}' --json
```
For any server command not yet mapped to a CLI subcommand.
---
## Error Handling
| Error | Likely Cause | Fix |
|-------|-------------|-----|
| `npx neuroskill status` hangs | NeuroSkill app not running | Open NeuroSkill desktop app |
| `device.state: "disconnected"` | BCI device not connected | Check Bluetooth, device battery |
| All scores return 0 | Poor electrode contact | Reposition headband, moisten electrodes |
| `signal_quality` values &lt; 0.7 | Loose electrodes | Adjust fit, clean electrode contacts |
| SNR &lt; 3 dB | Noisy signal | Minimize head movement, check environment |
| `command not found: npx` | Node.js not installed | Install Node.js 20+ |
---
## Example Interactions
**"How am I doing right now?"**
```bash
npx neuroskill status --json
```
→ Interpret scores naturally, mentioning focus, relaxation, mood, and any notable
ratios (FAA, TBR). Suggest an action only if metrics indicate a need.
**"I can't concentrate"**
```bash
npx neuroskill status --json
```
→ Check if metrics confirm it (high theta, low beta, rising TBR, high drowsiness).
→ If confirmed, suggest an appropriate protocol from `references/protocols.md`.
→ If metrics look fine, the issue may be motivational rather than neurological.
**"Compare my focus today vs yesterday"**
```bash
npx neuroskill compare --json
```
→ Interpret trends, not just numbers. Mention what improved, what declined, and
possible causes.
**"When was I last in a flow state?"**
```bash
npx neuroskill search-labels "flow" --json
npx neuroskill search --json
```
→ Report timestamps, associated metrics, and what the user was doing (from labels).
**"How did I sleep?"**
```bash
npx neuroskill sleep --json
```
→ Report sleep architecture (N3%, REM%, efficiency), compare to healthy targets,
and note any issues (high wake epochs, low REM).
**"Mark this moment — I just had a breakthrough"**
```bash
npx neuroskill label "breakthrough"
```
→ Confirm label saved. Optionally note the current metrics to remember the state.
---
## References
- [NeuroSkill Paper — arXiv:2603.03212](https://arxiv.org/abs/2603.03212) (Kosmyna & Hauptmann, MIT Media Lab)
- [NeuroSkill Desktop App](https://github.com/NeuroSkill-com/skill) (GPLv3)
- [NeuroLoop CLI Companion](https://github.com/NeuroSkill-com/neuroloop) (GPLv3)
- [MIT Media Lab Project](https://www.media.mit.edu/projects/neuroskill/overview/)

314
website/docs/user-guide/skills/optional/mcp/mcp-fastmcp.md Normal file
View file

@ -0,0 +1,314 @@
---
title: "Fastmcp — Build, test, inspect, install, and deploy MCP servers with FastMCP in Python"
sidebar_label: "Fastmcp"
description: "Build, test, inspect, install, and deploy MCP servers with FastMCP in Python"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Fastmcp
Build, test, inspect, install, and deploy MCP servers with FastMCP in Python. Use when creating a new MCP server, wrapping an API or database as MCP tools, exposing resources or prompts, or preparing a FastMCP server for Claude Code, Cursor, or HTTP deployment.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mcp/fastmcp` |
| Path | `optional-skills/mcp/fastmcp` |
| Version | `1.0.0` |
| Author | Hermes Agent |
| License | MIT |
| Tags | `MCP`, `FastMCP`, `Python`, `Tools`, `Resources`, `Prompts`, `Deployment` |
| Related skills | [`native-mcp`](/docs/user-guide/skills/bundled/mcp/mcp-native-mcp), [`mcporter`](/docs/user-guide/skills/optional/mcp/mcp-mcporter) |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# FastMCP
Build MCP servers in Python with FastMCP, validate them locally, install them into MCP clients, and deploy them as HTTP endpoints.
## When to Use
Use this skill when the task is to:
- create a new MCP server in Python
- wrap an API, database, CLI, or file-processing workflow as MCP tools
- expose resources or prompts in addition to tools
- smoke-test a server with the FastMCP CLI before wiring it into Hermes or another client
- install a server into Claude Code, Claude Desktop, Cursor, or a similar MCP client
- prepare a FastMCP server repo for HTTP deployment
Use `native-mcp` when the server already exists and only needs to be connected to Hermes. Use `mcporter` when the goal is ad-hoc CLI access to an existing MCP server instead of building one.
## Prerequisites
Install FastMCP in the working environment first:
```bash
pip install fastmcp
fastmcp version
```
For the API template, install `httpx` if it is not already present:
```bash
pip install httpx
```
## Included Files
### Templates
- `templates/api_wrapper.py` - REST API wrapper with auth header support
- `templates/database_server.py` - read-only SQLite query server
- `templates/file_processor.py` - text-file inspection and search server
### Scripts
- `scripts/scaffold_fastmcp.py` - copy a starter template and replace the server name placeholder
### References
- `references/fastmcp-cli.md` - FastMCP CLI workflow, installation targets, and deployment checks
## Workflow
### 1. Pick the Smallest Viable Server Shape
Choose the narrowest useful surface area first:
- API wrapper: start with 1-3 high-value endpoints, not the whole API
- database server: expose read-only introspection and a constrained query path
- file processor: expose deterministic operations with explicit path arguments
- prompts/resources: add only when the client needs reusable prompt templates or discoverable documents
Prefer a thin server with good names, docstrings, and schemas over a large server with vague tools.
### 2. Scaffold from a Template
Copy a template directly or use the scaffold helper:
```bash
python ~/.hermes/skills/mcp/fastmcp/scripts/scaffold_fastmcp.py \
--template api_wrapper \
--name "Acme API" \
--output ./acme_server.py
```
Available templates:
```bash
python ~/.hermes/skills/mcp/fastmcp/scripts/scaffold_fastmcp.py --list
```
If copying manually, replace `__SERVER_NAME__` with a real server name.
### 3. Implement Tools First
Start with `@mcp.tool` functions before adding resources or prompts.
Rules for tool design:
- Give every tool a concrete verb-based name
- Write docstrings as user-facing tool descriptions
- Keep parameters explicit and typed
- Return structured JSON-safe data where possible
- Validate unsafe inputs early
- Prefer read-only behavior by default for first versions
Good tool examples:
- `get_customer`
- `search_tickets`
- `describe_table`
- `summarize_text_file`
Weak tool examples:
- `run`
- `process`
- `do_thing`
### 4. Add Resources and Prompts Only When They Help
Add `@mcp.resource` when the client benefits from fetching stable read-only content such as schemas, policy docs, or generated reports.
Add `@mcp.prompt` when the server should provide a reusable prompt template for a known workflow.
Do not turn every document into a prompt. Prefer:
- tools for actions
- resources for data/document retrieval
- prompts for reusable LLM instructions
### 5. Test the Server Before Integrating It Anywhere
Use the FastMCP CLI for local validation:
```bash
fastmcp inspect acme_server.py:mcp
fastmcp list acme_server.py --json
fastmcp call acme_server.py search_resources query=router limit=5 --json
```
For fast iterative debugging, run the server locally:
```bash
fastmcp run acme_server.py:mcp
```
To test HTTP transport locally:
```bash
fastmcp run acme_server.py:mcp --transport http --host 127.0.0.1 --port 8000
fastmcp list http://127.0.0.1:8000/mcp --json
fastmcp call http://127.0.0.1:8000/mcp search_resources query=router --json
```
Always run at least one real `fastmcp call` against each new tool before claiming the server works.
### 6. Install into a Client When Local Validation Passes
FastMCP can register the server with supported MCP clients:
```bash
fastmcp install claude-code acme_server.py
fastmcp install claude-desktop acme_server.py
fastmcp install cursor acme_server.py -e .
```
Use `fastmcp discover` to inspect named MCP servers already configured on the machine.
When the goal is Hermes integration, either:
- configure the server in `~/.hermes/config.yaml` using the `native-mcp` skill, or
- keep using FastMCP CLI commands during development until the interface stabilizes
### 7. Deploy After the Local Contract Is Stable
For managed hosting, Prefect Horizon is the path FastMCP documents most directly. Before deployment:
```bash
fastmcp inspect acme_server.py:mcp
```
Make sure the repo contains:
- a Python file with the FastMCP server object
- `requirements.txt` or `pyproject.toml`
- any environment-variable documentation needed for deployment
For generic HTTP hosting, validate the HTTP transport locally first, then deploy on any Python-compatible platform that can expose the server port.
## Common Patterns
### API Wrapper Pattern
Use when exposing a REST or HTTP API as MCP tools.
Recommended first slice:
- one read path
- one list/search path
- optional health check
Implementation notes:
- keep auth in environment variables, not hardcoded
- centralize request logic in one helper
- surface API errors with concise context
- normalize inconsistent upstream payloads before returning them
Start from `templates/api_wrapper.py`.
### Database Pattern
Use when exposing safe query and inspection capabilities.
Recommended first slice:
- `list_tables`
- `describe_table`
- one constrained read query tool
Implementation notes:
- default to read-only DB access
- reject non-`SELECT` SQL in early versions
- limit row counts
- return rows plus column names
Start from `templates/database_server.py`.
### File Processor Pattern
Use when the server needs to inspect or transform files on demand.
Recommended first slice:
- summarize file contents
- search within files
- extract deterministic metadata
Implementation notes:
- accept explicit file paths
- check for missing files and encoding failures
- cap previews and result counts
- avoid shelling out unless a specific external tool is required
Start from `templates/file_processor.py`.
## Quality Bar
Before handing off a FastMCP server, verify all of the following:
- server imports cleanly
- `fastmcp inspect <file.py:mcp>` succeeds
- `fastmcp list <server spec> --json` succeeds
- every new tool has at least one real `fastmcp call`
- environment variables are documented
- the tool surface is small enough to understand without guesswork
## Troubleshooting
### FastMCP command missing
Install the package in the active environment:
```bash
pip install fastmcp
fastmcp version
```
### `fastmcp inspect` fails
Check that:
- the file imports without side effects that crash
- the FastMCP instance is named correctly in `<file.py:object>`
- optional dependencies from the template are installed
### Tool works in Python but not through CLI
Run:
```bash
fastmcp list server.py --json
fastmcp call server.py your_tool_name --json
```
This usually exposes naming mismatches, missing required arguments, or non-serializable return values.
### Hermes cannot see the deployed server
The server-building part may be correct while the Hermes config is not. Load the `native-mcp` skill and configure the server in `~/.hermes/config.yaml`, then restart Hermes.
## References
For CLI details, install targets, and deployment checks, read `references/fastmcp-cli.md`.

137
website/docs/user-guide/skills/optional/mcp/mcp-mcporter.md Normal file
View file

@ -0,0 +1,137 @@
---
title: "Mcporter"
sidebar_label: "Mcporter"
description: "Use the mcporter CLI to list, configure, auth, and call MCP servers/tools directly (HTTP or stdio), including ad-hoc servers, config edits, and CLI/type gene..."
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Mcporter
Use the mcporter CLI to list, configure, auth, and call MCP servers/tools directly (HTTP or stdio), including ad-hoc servers, config edits, and CLI/type generation.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mcp/mcporter` |
| Path | `optional-skills/mcp/mcporter` |
| Version | `1.0.0` |
| Author | community |
| License | MIT |
| Tags | `MCP`, `Tools`, `API`, `Integrations`, `Interop` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# mcporter
Use `mcporter` to discover, call, and manage [MCP (Model Context Protocol)](https://modelcontextprotocol.io/) servers and tools directly from the terminal.
## Prerequisites
Requires Node.js:
```bash
# No install needed (runs via npx)
npx mcporter list
# Or install globally
npm install -g mcporter
```
## Quick Start
```bash
# List MCP servers already configured on this machine
mcporter list
# List tools for a specific server with schema details
mcporter list <server> --schema
# Call a tool
mcporter call <server.tool> key=value
```
## Discovering MCP Servers
mcporter auto-discovers servers configured by other MCP clients (Claude Desktop, Cursor, etc.) on the machine. To find new servers to use, browse registries like [mcpfinder.dev](https://mcpfinder.dev) or [mcp.so](https://mcp.so), then connect ad-hoc:
```bash
# Connect to any MCP server by URL (no config needed)
mcporter list --http-url https://some-mcp-server.com --name my_server
# Or run a stdio server on the fly
mcporter list --stdio "npx -y @modelcontextprotocol/server-filesystem" --name fs
```
## Calling Tools
```bash
# Key=value syntax
mcporter call linear.list_issues team=ENG limit:5
# Function syntax
mcporter call "linear.create_issue(title: \"Bug fix needed\")"
# Ad-hoc HTTP server (no config needed)
mcporter call https://api.example.com/mcp.fetch url=https://example.com
# Ad-hoc stdio server
mcporter call --stdio "bun run ./server.ts" scrape url=https://example.com
# JSON payload
mcporter call <server.tool> --args '{"limit": 5}'
# Machine-readable output (recommended for Hermes)
mcporter call <server.tool> key=value --output json
```
## Auth and Config
```bash
# OAuth login for a server
mcporter auth <server | url> [--reset]
# Manage config
mcporter config list
mcporter config get <key>
mcporter config add <server>
mcporter config remove <server>
mcporter config import <path>
```
Config file location: `./config/mcporter.json` (override with `--config`).
## Daemon
For persistent server connections:
```bash
mcporter daemon start
mcporter daemon status
mcporter daemon stop
mcporter daemon restart
```
## Code Generation
```bash
# Generate a CLI wrapper for an MCP server
mcporter generate-cli --server <name>
mcporter generate-cli --command <url>
# Inspect a generated CLI
mcporter inspect-cli <path> [--json]
# Generate TypeScript types/client
mcporter emit-ts <server> --mode client
mcporter emit-ts <server> --mode types
```
## Notes
- Use `--output json` for structured output that's easier to parse
- Ad-hoc servers (HTTP URL or `--stdio` command) work without any config — useful for one-off calls
- OAuth auth may require interactive browser flow — use `terminal(command="mcporter auth <server>", pty=true)` if needed

315
website/docs/user-guide/skills/optional/migration/migration-openclaw-migration.md Normal file
View file

@ -0,0 +1,315 @@
---
title: "Openclaw Migration — Migrate a user's OpenClaw customization footprint into Hermes Agent"
sidebar_label: "Openclaw Migration"
description: "Migrate a user's OpenClaw customization footprint into Hermes Agent"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Openclaw Migration
Migrate a user's OpenClaw customization footprint into Hermes Agent. Imports Hermes-compatible memories, SOUL.md, command allowlists, user skills, and selected workspace assets from ~/.openclaw, then reports exactly what could not be migrated and why.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/migration/openclaw-migration` |
| Path | `optional-skills/migration/openclaw-migration` |
| Version | `1.0.0` |
| Author | Hermes Agent (Nous Research) |
| License | MIT |
| Tags | `Migration`, `OpenClaw`, `Hermes`, `Memory`, `Persona`, `Import` |
| Related skills | [`hermes-agent`](/docs/user-guide/skills/bundled/autonomous-ai-agents/autonomous-ai-agents-hermes-agent) |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# OpenClaw -> Hermes Migration
Use this skill when a user wants to move their OpenClaw setup into Hermes Agent with minimal manual cleanup.
## CLI Command
For a quick, non-interactive migration, use the built-in CLI command:
```bash
hermes claw migrate # Full interactive migration
hermes claw migrate --dry-run # Preview what would be migrated
hermes claw migrate --preset user-data # Migrate without secrets
hermes claw migrate --overwrite # Overwrite existing conflicts
hermes claw migrate --source /custom/path/.openclaw # Custom source
```
The CLI command runs the same migration script described below. Use this skill (via the agent) when you want an interactive, guided migration with dry-run previews and per-item conflict resolution.
**First-time setup:** The `hermes setup` wizard automatically detects `~/.openclaw` and offers migration before configuration begins.
## What this skill does
It uses `scripts/openclaw_to_hermes.py` to:
- import `SOUL.md` into the Hermes home directory as `SOUL.md`
- transform OpenClaw `MEMORY.md` and `USER.md` into Hermes memory entries
- merge OpenClaw command approval patterns into Hermes `command_allowlist`
- migrate Hermes-compatible messaging settings such as `TELEGRAM_ALLOWED_USERS` and `MESSAGING_CWD`
- copy OpenClaw skills into `~/.hermes/skills/openclaw-imports/`
- optionally copy the OpenClaw workspace instructions file into a chosen Hermes workspace
- mirror compatible workspace assets such as `workspace/tts/` into `~/.hermes/tts/`
- archive non-secret docs that do not have a direct Hermes destination
- produce a structured report listing migrated items, conflicts, skipped items, and reasons
## Path resolution
The helper script lives in this skill directory at:
- `scripts/openclaw_to_hermes.py`
When this skill is installed from the Skills Hub, the normal location is:
- `~/.hermes/skills/migration/openclaw-migration/scripts/openclaw_to_hermes.py`
Do not guess a shorter path like `~/.hermes/skills/openclaw-migration/...`.
Before running the helper:
1. Prefer the installed path under `~/.hermes/skills/migration/openclaw-migration/`.
2. If that path fails, inspect the installed skill directory and resolve the script relative to the installed `SKILL.md`.
3. Only use `find` as a fallback if the installed location is missing or the skill was moved manually.
4. When calling the terminal tool, do not pass `workdir: "~"`. Use an absolute directory such as the user's home directory, or omit `workdir` entirely.
With `--migrate-secrets`, it will also import a small allowlisted set of Hermes-compatible secrets, currently:
- `TELEGRAM_BOT_TOKEN`
## Default workflow
1. Inspect first with a dry run.
2. Present a simple summary of what can be migrated, what cannot be migrated, and what would be archived.
3. If the `clarify` tool is available, use it for user decisions instead of asking for a free-form prose reply.
4. If the dry run finds imported skill directory conflicts, ask how those should be handled before executing.
5. Ask the user to choose between the two supported migration modes before executing.
6. Ask for a target workspace path only if the user wants the workspace instructions file brought over.
7. Execute the migration with the matching preset and flags.
8. Summarize the results, especially:
- what was migrated
- what was archived for manual review
- what was skipped and why
## User interaction protocol
Hermes CLI supports the `clarify` tool for interactive prompts, but it is limited to:
- one choice at a time
- up to 4 predefined choices
- an automatic `Other` free-text option
It does **not** support true multi-select checkboxes in a single prompt.
For every `clarify` call:
- always include a non-empty `question`
- include `choices` only for real selectable prompts
- keep `choices` to 2-4 plain string options
- never emit placeholder or truncated options such as `...`
- never pad or stylize choices with extra whitespace
- never include fake form fields in the question such as `enter directory here`, blank lines to fill in, or underscores like `_____`
- for open-ended path questions, ask only the plain sentence; the user types in the normal CLI prompt below the panel
If a `clarify` call returns an error, inspect the error text, correct the payload, and retry once with a valid `question` and clean choices.
When `clarify` is available and the dry run reveals any required user decision, your **next action must be a `clarify` tool call**.
Do not end the turn with a normal assistant message such as:
- "Let me present the choices"
- "What would you like to do?"
- "Here are the options"
If a user decision is required, collect it via `clarify` before producing more prose.
If multiple unresolved decisions remain, do not insert an explanatory assistant message between them. After one `clarify` response is received, your next action should usually be the next required `clarify` call.
Treat `workspace-agents` as an unresolved decision whenever the dry run reports:
- `kind="workspace-agents"`
- `status="skipped"`
- reason containing `No workspace target was provided`
In that case, you must ask about workspace instructions before execution. Do not silently treat that as a decision to skip.
Because of that limitation, use this simplified decision flow:
1. For `SOUL.md` conflicts, use `clarify` with choices such as:
- `keep existing`
- `overwrite with backup`
- `review first`
2. If the dry run shows one or more `kind="skill"` items with `status="conflict"`, use `clarify` with choices such as:
- `keep existing skills`
- `overwrite conflicting skills with backup`
- `import conflicting skills under renamed folders`
3. For workspace instructions, use `clarify` with choices such as:
- `skip workspace instructions`
- `copy to a workspace path`
- `decide later`
4. If the user chooses to copy workspace instructions, ask a follow-up open-ended `clarify` question requesting an **absolute path**.
5. If the user chooses `skip workspace instructions` or `decide later`, proceed without `--workspace-target`.
5. For migration mode, use `clarify` with these 3 choices:
- `user-data only`
- `full compatible migration`
- `cancel`
6. `user-data only` means: migrate user data and compatible config, but do **not** import allowlisted secrets.
7. `full compatible migration` means: migrate the same compatible user data plus the allowlisted secrets when present.
8. If `clarify` is not available, ask the same question in normal text, but still constrain the answer to `user-data only`, `full compatible migration`, or `cancel`.
Execution gate:
- Do not execute while a `workspace-agents` skip caused by `No workspace target was provided` remains unresolved.
- The only valid ways to resolve it are:
- user explicitly chooses `skip workspace instructions`
- user explicitly chooses `decide later`
- user provides a workspace path after choosing `copy to a workspace path`
- Absence of a workspace target in the dry run is not itself permission to execute.
- Do not execute while any required `clarify` decision remains unresolved.
Use these exact `clarify` payload shapes as the default pattern:
- `{"question":"Your existing SOUL.md conflicts with the imported one. What should I do?","choices":["keep existing","overwrite with backup","review first"]}`
- `{"question":"One or more imported OpenClaw skills already exist in Hermes. How should I handle those skill conflicts?","choices":["keep existing skills","overwrite conflicting skills with backup","import conflicting skills under renamed folders"]}`
- `{"question":"Choose migration mode: migrate only user data, or run the full compatible migration including allowlisted secrets?","choices":["user-data only","full compatible migration","cancel"]}`
- `{"question":"Do you want to copy the OpenClaw workspace instructions file into a Hermes workspace?","choices":["skip workspace instructions","copy to a workspace path","decide later"]}`
- `{"question":"Please provide an absolute path where the workspace instructions should be copied."}`
## Decision-to-command mapping
Map user decisions to command flags exactly:
- If the user chooses `keep existing` for `SOUL.md`, do **not** add `--overwrite`.
- If the user chooses `overwrite with backup`, add `--overwrite`.
- If the user chooses `review first`, stop before execution and review the relevant files.
- If the user chooses `keep existing skills`, add `--skill-conflict skip`.
- If the user chooses `overwrite conflicting skills with backup`, add `--skill-conflict overwrite`.
- If the user chooses `import conflicting skills under renamed folders`, add `--skill-conflict rename`.
- If the user chooses `user-data only`, execute with `--preset user-data` and do **not** add `--migrate-secrets`.
- If the user chooses `full compatible migration`, execute with `--preset full --migrate-secrets`.
- Only add `--workspace-target` if the user explicitly provided an absolute workspace path.
- If the user chooses `skip workspace instructions` or `decide later`, do not add `--workspace-target`.
Before executing, restate the exact command plan in plain language and make sure it matches the user's choices.
## Post-run reporting rules
After execution, treat the script's JSON output as the source of truth.
1. Base all counts on `report.summary`.
2. Only list an item under "Successfully Migrated" if its `status` is exactly `migrated`.
3. Do not claim a conflict was resolved unless the report shows that item as `migrated`.
4. Do not say `SOUL.md` was overwritten unless the report item for `kind="soul"` has `status="migrated"`.
5. If `report.summary.conflict > 0`, include a conflict section instead of silently implying success.
6. If counts and listed items disagree, fix the list to match the report before responding.
7. Include the `output_dir` path from the report when available so the user can inspect `report.json`, `summary.md`, backups, and archived files.
8. For memory or user-profile overflow, do not say the entries were archived unless the report explicitly shows an archive path. If `details.overflow_file` exists, say the full overflow list was exported there.
9. If a skill was imported under a renamed folder, report the final destination and mention `details.renamed_from`.
10. If `report.skill_conflict_mode` is present, use it as the source of truth for the selected imported-skill conflict policy.
11. If an item has `status="skipped"`, do not describe it as overwritten, backed up, migrated, or resolved.
12. If `kind="soul"` has `status="skipped"` with reason `Target already matches source`, say it was left unchanged and do not mention a backup.
13. If a renamed imported skill has an empty `details.backup`, do not imply the existing Hermes skill was renamed or backed up. Say only that the imported copy was placed in the new destination and reference `details.renamed_from` as the pre-existing folder that remained in place.
## Migration presets
Prefer these two presets in normal use:
- `user-data`
- `full`
`user-data` includes:
- `soul`
- `workspace-agents`
- `memory`
- `user-profile`
- `messaging-settings`
- `command-allowlist`
- `skills`
- `tts-assets`
- `archive`
`full` includes everything in `user-data` plus:
- `secret-settings`
The helper script still supports category-level `--include` / `--exclude`, but treat that as an advanced fallback rather than the default UX.
## Commands
Dry run with full discovery:
```bash
python3 ~/.hermes/skills/migration/openclaw-migration/scripts/openclaw_to_hermes.py
```
When using the terminal tool, prefer an absolute invocation pattern such as:
```json
{"command":"python3 /home/USER/.hermes/skills/migration/openclaw-migration/scripts/openclaw_to_hermes.py","workdir":"/home/USER"}
```
Dry run with the user-data preset:
```bash
python3 ~/.hermes/skills/migration/openclaw-migration/scripts/openclaw_to_hermes.py --preset user-data
```
Execute a user-data migration:
```bash
python3 ~/.hermes/skills/migration/openclaw-migration/scripts/openclaw_to_hermes.py --execute --preset user-data --skill-conflict skip
```
Execute a full compatible migration:
```bash
python3 ~/.hermes/skills/migration/openclaw-migration/scripts/openclaw_to_hermes.py --execute --preset full --migrate-secrets --skill-conflict skip
```
Execute with workspace instructions included:
```bash
python3 ~/.hermes/skills/migration/openclaw-migration/scripts/openclaw_to_hermes.py --execute --preset user-data --skill-conflict rename --workspace-target "/absolute/workspace/path"
```
Do not use `$PWD` or the home directory as the workspace target by default. Ask for an explicit workspace path first.
## Important rules
1. Run a dry run before writing unless the user explicitly says to proceed immediately.
2. Do not migrate secrets by default. Tokens, auth blobs, device credentials, and raw gateway config should stay out of Hermes unless the user explicitly asks for secret migration.
3. Do not silently overwrite non-empty Hermes targets unless the user explicitly wants that. The helper script will preserve backups when overwriting is enabled.
4. Always give the user the skipped-items report. That report is part of the migration, not an optional extra.
5. Prefer the primary OpenClaw workspace (`~/.openclaw/workspace/`) over `workspace.default/`. Only use the default workspace as fallback when the primary files are missing.
6. Even in secret-migration mode, only migrate secrets with a clean Hermes destination. Unsupported auth blobs must still be reported as skipped.
7. If the dry run shows a large asset copy, a conflicting `SOUL.md`, or overflowed memory entries, call those out separately before execution.
8. Default to `user-data only` if the user is unsure.
9. Only include `workspace-agents` when the user has explicitly provided a destination workspace path.
10. Treat category-level `--include` / `--exclude` as an advanced escape hatch, not the normal flow.
11. Do not end the dry-run summary with a vague “What would you like to do?” if `clarify` is available. Use structured follow-up prompts instead.
12. Do not use an open-ended `clarify` prompt when a real choice prompt would work. Prefer selectable choices first, then free text only for absolute paths or file review requests.
13. After a dry run, never stop after summarizing if there is still an unresolved decision. Use `clarify` immediately for the highest-priority blocking decision.
14. Priority order for follow-up questions:
- `SOUL.md` conflict
- imported skill conflicts
- migration mode
- workspace instructions destination
15. Do not promise to present choices later in the same message. Present them by actually calling `clarify`.
16. After the migration-mode answer, explicitly check whether `workspace-agents` is still unresolved. If it is, your next action must be the workspace-instructions `clarify` call.
17. After any `clarify` answer, if another required decision remains, do not narrate what was just decided. Ask the next required question immediately.
## Expected result
After a successful run, the user should have:
- Hermes persona state imported
- Hermes memory files populated with converted OpenClaw knowledge
- OpenClaw skills available under `~/.hermes/skills/openclaw-imports/`
- a migration report showing any conflicts, omissions, or unsupported data

349
website/docs/user-guide/skills/optional/mlops/mlops-accelerate.md Normal file
View file

@ -0,0 +1,349 @@
---
title: "Huggingface Accelerate — Simplest distributed training API"
sidebar_label: "Huggingface Accelerate"
description: "Simplest distributed training API"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Huggingface Accelerate
Simplest distributed training API. 4 lines to add distributed support to any PyTorch script. Unified API for DeepSpeed/FSDP/Megatron/DDP. Automatic device placement, mixed precision (FP16/BF16/FP8). Interactive config, single launch command. HuggingFace ecosystem standard.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/accelerate` |
| Path | `optional-skills/mlops/accelerate` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `accelerate`, `torch`, `transformers` |
| Tags | `Distributed Training`, `HuggingFace`, `Accelerate`, `DeepSpeed`, `FSDP`, `Mixed Precision`, `PyTorch`, `DDP`, `Unified API`, `Simple` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# HuggingFace Accelerate - Unified Distributed Training
## Quick start
Accelerate simplifies distributed training to 4 lines of code.
**Installation**:
```bash
pip install accelerate
```
**Convert PyTorch script** (4 lines):
```python
import torch
+ from accelerate import Accelerator
+ accelerator = Accelerator()
model = torch.nn.Transformer()
optimizer = torch.optim.Adam(model.parameters())
dataloader = torch.utils.data.DataLoader(dataset)
+ model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
for batch in dataloader:
optimizer.zero_grad()
loss = model(batch)
- loss.backward()
+ accelerator.backward(loss)
optimizer.step()
```
**Run** (single command):
```bash
accelerate launch train.py
```
## Common workflows
### Workflow 1: From single GPU to multi-GPU
**Original script**:
```python
# train.py
import torch
model = torch.nn.Linear(10, 2).to('cuda')
optimizer = torch.optim.Adam(model.parameters())
dataloader = torch.utils.data.DataLoader(dataset, batch_size=32)
for epoch in range(10):
for batch in dataloader:
batch = batch.to('cuda')
optimizer.zero_grad()
loss = model(batch).mean()
loss.backward()
optimizer.step()
```
**With Accelerate** (4 lines added):
```python
# train.py
import torch
from accelerate import Accelerator # +1
accelerator = Accelerator() # +2
model = torch.nn.Linear(10, 2)
optimizer = torch.optim.Adam(model.parameters())
dataloader = torch.utils.data.DataLoader(dataset, batch_size=32)
model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader) # +3
for epoch in range(10):
for batch in dataloader:
# No .to('cuda') needed - automatic!
optimizer.zero_grad()
loss = model(batch).mean()
accelerator.backward(loss) # +4
optimizer.step()
```
**Configure** (interactive):
```bash
accelerate config
```
**Questions**:
- Which machine? (single/multi GPU/TPU/CPU)
- How many machines? (1)
- Mixed precision? (no/fp16/bf16/fp8)
- DeepSpeed? (no/yes)
**Launch** (works on any setup):
```bash
# Single GPU
accelerate launch train.py
# Multi-GPU (8 GPUs)
accelerate launch --multi_gpu --num_processes 8 train.py
# Multi-node
accelerate launch --multi_gpu --num_processes 16 \
--num_machines 2 --machine_rank 0 \
--main_process_ip $MASTER_ADDR \
train.py
```
### Workflow 2: Mixed precision training
**Enable FP16/BF16**:
```python
from accelerate import Accelerator
# FP16 (with gradient scaling)
accelerator = Accelerator(mixed_precision='fp16')
# BF16 (no scaling, more stable)
accelerator = Accelerator(mixed_precision='bf16')
# FP8 (H100+)
accelerator = Accelerator(mixed_precision='fp8')
model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
# Everything else is automatic!
for batch in dataloader:
with accelerator.autocast(): # Optional, done automatically
loss = model(batch)
accelerator.backward(loss)
```
### Workflow 3: DeepSpeed ZeRO integration
**Enable DeepSpeed ZeRO-2**:
```python
from accelerate import Accelerator
accelerator = Accelerator(
mixed_precision='bf16',
deepspeed_plugin={
"zero_stage": 2, # ZeRO-2
"offload_optimizer": False,
"gradient_accumulation_steps": 4
}
)
# Same code as before!
model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
```
**Or via config**:
```bash
accelerate config
# Select: DeepSpeed → ZeRO-2
```
**deepspeed_config.json**:
```json
{
"fp16": {"enabled": false},
"bf16": {"enabled": true},
"zero_optimization": {
"stage": 2,
"offload_optimizer": {"device": "cpu"},
"allgather_bucket_size": 5e8,
"reduce_bucket_size": 5e8
}
}
```
**Launch**:
```bash
accelerate launch --config_file deepspeed_config.json train.py
```
### Workflow 4: FSDP (Fully Sharded Data Parallel)
**Enable FSDP**:
```python
from accelerate import Accelerator, FullyShardedDataParallelPlugin
fsdp_plugin = FullyShardedDataParallelPlugin(
sharding_strategy="FULL_SHARD", # ZeRO-3 equivalent
auto_wrap_policy="TRANSFORMER_AUTO_WRAP",
cpu_offload=False
)
accelerator = Accelerator(
mixed_precision='bf16',
fsdp_plugin=fsdp_plugin
)
model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
```
**Or via config**:
```bash
accelerate config
# Select: FSDP → Full Shard → No CPU Offload
```
### Workflow 5: Gradient accumulation
**Accumulate gradients**:
```python
from accelerate import Accelerator
accelerator = Accelerator(gradient_accumulation_steps=4)
model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
for batch in dataloader:
with accelerator.accumulate(model): # Handles accumulation
optimizer.zero_grad()
loss = model(batch)
accelerator.backward(loss)
optimizer.step()
```
**Effective batch size**: `batch_size * num_gpus * gradient_accumulation_steps`
## When to use vs alternatives
**Use Accelerate when**:
- Want simplest distributed training
- Need single script for any hardware
- Use HuggingFace ecosystem
- Want flexibility (DDP/DeepSpeed/FSDP/Megatron)
- Need quick prototyping
**Key advantages**:
- **4 lines**: Minimal code changes
- **Unified API**: Same code for DDP, DeepSpeed, FSDP, Megatron
- **Automatic**: Device placement, mixed precision, sharding
- **Interactive config**: No manual launcher setup
- **Single launch**: Works everywhere
**Use alternatives instead**:
- **PyTorch Lightning**: Need callbacks, high-level abstractions
- **Ray Train**: Multi-node orchestration, hyperparameter tuning
- **DeepSpeed**: Direct API control, advanced features
- **Raw DDP**: Maximum control, minimal abstraction
## Common issues
**Issue: Wrong device placement**
Don't manually move to device:
```python
# WRONG
batch = batch.to('cuda')
# CORRECT
# Accelerate handles it automatically after prepare()
```
**Issue: Gradient accumulation not working**
Use context manager:
```python
# CORRECT
with accelerator.accumulate(model):
optimizer.zero_grad()
accelerator.backward(loss)
optimizer.step()
```
**Issue: Checkpointing in distributed**
Use accelerator methods:
```python
# Save only on main process
if accelerator.is_main_process:
accelerator.save_state('checkpoint/')
# Load on all processes
accelerator.load_state('checkpoint/')
```
**Issue: Different results with FSDP**
Ensure same random seed:
```python
from accelerate.utils import set_seed
set_seed(42)
```
## Advanced topics
**Megatron integration**: See [references/megatron-integration.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/accelerate/references/megatron-integration.md) for tensor parallelism, pipeline parallelism, and sequence parallelism setup.
**Custom plugins**: See [references/custom-plugins.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/accelerate/references/custom-plugins.md) for creating custom distributed plugins and advanced configuration.
**Performance tuning**: See [references/performance.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/accelerate/references/performance.md) for profiling, memory optimization, and best practices.
## Hardware requirements
- **CPU**: Works (slow)
- **Single GPU**: Works
- **Multi-GPU**: DDP (default), DeepSpeed, or FSDP
- **Multi-node**: DDP, DeepSpeed, FSDP, Megatron
- **TPU**: Supported
- **Apple MPS**: Supported
**Launcher requirements**:
- **DDP**: `torch.distributed.run` (built-in)
- **DeepSpeed**: `deepspeed` (pip install deepspeed)
- **FSDP**: PyTorch 1.12+ (built-in)
- **Megatron**: Custom setup
## Resources
- Docs: https://huggingface.co/docs/accelerate
- GitHub: https://github.com/huggingface/accelerate
- Version: 1.11.0+
- Tutorial: "Accelerate your scripts"
- Examples: https://github.com/huggingface/accelerate/tree/main/examples
- Used by: HuggingFace Transformers, TRL, PEFT, all HF libraries

424
website/docs/user-guide/skills/optional/mlops/mlops-chroma.md Normal file
View file

@ -0,0 +1,424 @@
---
title: "Chroma — Open-source embedding database for AI applications"
sidebar_label: "Chroma"
description: "Open-source embedding database for AI applications"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Chroma
Open-source embedding database for AI applications. Store embeddings and metadata, perform vector and full-text search, filter by metadata. Simple 4-function API. Scales from notebooks to production clusters. Use for semantic search, RAG applications, or document retrieval. Best for local development and open-source projects.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/chroma` |
| Path | `optional-skills/mlops/chroma` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `chromadb`, `sentence-transformers` |
| Tags | `RAG`, `Chroma`, `Vector Database`, `Embeddings`, `Semantic Search`, `Open Source`, `Self-Hosted`, `Document Retrieval`, `Metadata Filtering` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Chroma - Open-Source Embedding Database
The AI-native database for building LLM applications with memory.
## When to use Chroma
**Use Chroma when:**
- Building RAG (retrieval-augmented generation) applications
- Need local/self-hosted vector database
- Want open-source solution (Apache 2.0)
- Prototyping in notebooks
- Semantic search over documents
- Storing embeddings with metadata
**Metrics**:
- **24,300+ GitHub stars**
- **1,900+ forks**
- **v1.3.3** (stable, weekly releases)
- **Apache 2.0 license**
**Use alternatives instead**:
- **Pinecone**: Managed cloud, auto-scaling
- **FAISS**: Pure similarity search, no metadata
- **Weaviate**: Production ML-native database
- **Qdrant**: High performance, Rust-based
## Quick start
### Installation
```bash
# Python
pip install chromadb
# JavaScript/TypeScript
npm install chromadb @chroma-core/default-embed
```
### Basic usage (Python)
```python
import chromadb
# Create client
client = chromadb.Client()
# Create collection
collection = client.create_collection(name="my_collection")
# Add documents
collection.add(
documents=["This is document 1", "This is document 2"],
metadatas=[{"source": "doc1"}, {"source": "doc2"}],
ids=["id1", "id2"]
)
# Query
results = collection.query(
query_texts=["document about topic"],
n_results=2
)
print(results)
```
## Core operations
### 1. Create collection
```python
# Simple collection
collection = client.create_collection("my_docs")
# With custom embedding function
from chromadb.utils import embedding_functions
openai_ef = embedding_functions.OpenAIEmbeddingFunction(
api_key="your-key",
model_name="text-embedding-3-small"
)
collection = client.create_collection(
name="my_docs",
embedding_function=openai_ef
)
# Get existing collection
collection = client.get_collection("my_docs")
# Delete collection
client.delete_collection("my_docs")
```
### 2. Add documents
```python
# Add with auto-generated IDs
collection.add(
documents=["Doc 1", "Doc 2", "Doc 3"],
metadatas=[
{"source": "web", "category": "tutorial"},
{"source": "pdf", "page": 5},
{"source": "api", "timestamp": "2025-01-01"}
],
ids=["id1", "id2", "id3"]
)
# Add with custom embeddings
collection.add(
embeddings=[[0.1, 0.2, ...], [0.3, 0.4, ...]],
documents=["Doc 1", "Doc 2"],
ids=["id1", "id2"]
)
```
### 3. Query (similarity search)
```python
# Basic query
results = collection.query(
query_texts=["machine learning tutorial"],
n_results=5
)
# Query with filters
results = collection.query(
query_texts=["Python programming"],
n_results=3,
where={"source": "web"}
)
# Query with metadata filters
results = collection.query(
query_texts=["advanced topics"],
where={
"$and": [
{"category": "tutorial"},
{"difficulty": {"$gte": 3}}
]
}
)
# Access results
print(results["documents"]) # List of matching documents
print(results["metadatas"]) # Metadata for each doc
print(results["distances"]) # Similarity scores
print(results["ids"]) # Document IDs
```
### 4. Get documents
```python
# Get by IDs
docs = collection.get(
ids=["id1", "id2"]
)
# Get with filters
docs = collection.get(
where={"category": "tutorial"},
limit=10
)
# Get all documents
docs = collection.get()
```
### 5. Update documents
```python
# Update document content
collection.update(
ids=["id1"],
documents=["Updated content"],
metadatas=[{"source": "updated"}]
)
```
### 6. Delete documents
```python
# Delete by IDs
collection.delete(ids=["id1", "id2"])
# Delete with filter
collection.delete(
where={"source": "outdated"}
)
```
## Persistent storage
```python
# Persist to disk
client = chromadb.PersistentClient(path="./chroma_db")
collection = client.create_collection("my_docs")
collection.add(documents=["Doc 1"], ids=["id1"])
# Data persisted automatically
# Reload later with same path
client = chromadb.PersistentClient(path="./chroma_db")
collection = client.get_collection("my_docs")
```
## Embedding functions
### Default (Sentence Transformers)
```python
# Uses sentence-transformers by default
collection = client.create_collection("my_docs")
# Default model: all-MiniLM-L6-v2
```
### OpenAI
```python
from chromadb.utils import embedding_functions
openai_ef = embedding_functions.OpenAIEmbeddingFunction(
api_key="your-key",
model_name="text-embedding-3-small"
)
collection = client.create_collection(
name="openai_docs",
embedding_function=openai_ef
)
```
### HuggingFace
```python
huggingface_ef = embedding_functions.HuggingFaceEmbeddingFunction(
api_key="your-key",
model_name="sentence-transformers/all-mpnet-base-v2"
)
collection = client.create_collection(
name="hf_docs",
embedding_function=huggingface_ef
)
```
### Custom embedding function
```python
from chromadb import Documents, EmbeddingFunction, Embeddings
class MyEmbeddingFunction(EmbeddingFunction):
def __call__(self, input: Documents) -> Embeddings:
# Your embedding logic
return embeddings
my_ef = MyEmbeddingFunction()
collection = client.create_collection(
name="custom_docs",
embedding_function=my_ef
)
```
## Metadata filtering
```python
# Exact match
results = collection.query(
query_texts=["query"],
where={"category": "tutorial"}
)
# Comparison operators
results = collection.query(
query_texts=["query"],
where={"page": {"$gt": 10}} # $gt, $gte, $lt, $lte, $ne
)
# Logical operators
results = collection.query(
query_texts=["query"],
where={
"$and": [
{"category": "tutorial"},
{"difficulty": {"$lte": 3}}
]
} # Also: $or
)
# Contains
results = collection.query(
query_texts=["query"],
where={"tags": {"$in": ["python", "ml"]}}
)
```
## LangChain integration
```python
from langchain_chroma import Chroma
from langchain_openai import OpenAIEmbeddings
from langchain.text_splitter import RecursiveCharacterTextSplitter
# Split documents
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000)
docs = text_splitter.split_documents(documents)
# Create Chroma vector store
vectorstore = Chroma.from_documents(
documents=docs,
embedding=OpenAIEmbeddings(),
persist_directory="./chroma_db"
)
# Query
results = vectorstore.similarity_search("machine learning", k=3)
# As retriever
retriever = vectorstore.as_retriever(search_kwargs={"k": 5})
```
## LlamaIndex integration
```python
from llama_index.vector_stores.chroma import ChromaVectorStore
from llama_index.core import VectorStoreIndex, StorageContext
import chromadb
# Initialize Chroma
db = chromadb.PersistentClient(path="./chroma_db")
collection = db.get_or_create_collection("my_collection")
# Create vector store
vector_store = ChromaVectorStore(chroma_collection=collection)
storage_context = StorageContext.from_defaults(vector_store=vector_store)
# Create index
index = VectorStoreIndex.from_documents(
documents,
storage_context=storage_context
)
# Query
query_engine = index.as_query_engine()
response = query_engine.query("What is machine learning?")
```
## Server mode
```python
# Run Chroma server
# Terminal: chroma run --path ./chroma_db --port 8000
# Connect to server
import chromadb
from chromadb.config import Settings
client = chromadb.HttpClient(
host="localhost",
port=8000,
settings=Settings(anonymized_telemetry=False)
)
# Use as normal
collection = client.get_or_create_collection("my_docs")
```
## Best practices
1. **Use persistent client** - Don't lose data on restart
2. **Add metadata** - Enables filtering and tracking
3. **Batch operations** - Add multiple docs at once
4. **Choose right embedding model** - Balance speed/quality
5. **Use filters** - Narrow search space
6. **Unique IDs** - Avoid collisions
7. **Regular backups** - Copy chroma_db directory
8. **Monitor collection size** - Scale up if needed
9. **Test embedding functions** - Ensure quality
10. **Use server mode for production** - Better for multi-user
## Performance
| Operation | Latency | Notes |
|-----------|---------|-------|
| Add 100 docs | ~1-3s | With embedding |
| Query (top 10) | ~50-200ms | Depends on collection size |
| Metadata filter | ~10-50ms | Fast with proper indexing |
## Resources
- **GitHub**: https://github.com/chroma-core/chroma ⭐ 24,300+
- **Docs**: https://docs.trychroma.com
- **Discord**: https://discord.gg/MMeYNTmh3x
- **Version**: 1.3.3+
- **License**: Apache 2.0

271
website/docs/user-guide/skills/optional/mlops/mlops-clip.md Normal file
View file

@ -0,0 +1,271 @@
---
title: "Clip — OpenAI's model connecting vision and language"
sidebar_label: "Clip"
description: "OpenAI's model connecting vision and language"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Clip
OpenAI's model connecting vision and language. Enables zero-shot image classification, image-text matching, and cross-modal retrieval. Trained on 400M image-text pairs. Use for image search, content moderation, or vision-language tasks without fine-tuning. Best for general-purpose image understanding.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/clip` |
| Path | `optional-skills/mlops/clip` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `transformers`, `torch`, `pillow` |
| Tags | `Multimodal`, `CLIP`, `Vision-Language`, `Zero-Shot`, `Image Classification`, `OpenAI`, `Image Search`, `Cross-Modal Retrieval`, `Content Moderation` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# CLIP - Contrastive Language-Image Pre-Training
OpenAI's model that understands images from natural language.
## When to use CLIP
**Use when:**
- Zero-shot image classification (no training data needed)
- Image-text similarity/matching
- Semantic image search
- Content moderation (detect NSFW, violence)
- Visual question answering
- Cross-modal retrieval (image→text, text→image)
**Metrics**:
- **25,300+ GitHub stars**
- Trained on 400M image-text pairs
- Matches ResNet-50 on ImageNet (zero-shot)
- MIT License
**Use alternatives instead**:
- **BLIP-2**: Better captioning
- **LLaVA**: Vision-language chat
- **Segment Anything**: Image segmentation
## Quick start
### Installation
```bash
pip install git+https://github.com/openai/CLIP.git
pip install torch torchvision ftfy regex tqdm
```
### Zero-shot classification
```python
import torch
import clip
from PIL import Image
# Load model
device = "cuda" if torch.cuda.is_available() else "cpu"
model, preprocess = clip.load("ViT-B/32", device=device)
# Load image
image = preprocess(Image.open("photo.jpg")).unsqueeze(0).to(device)
# Define possible labels
text = clip.tokenize(["a dog", "a cat", "a bird", "a car"]).to(device)
# Compute similarity
with torch.no_grad():
image_features = model.encode_image(image)
text_features = model.encode_text(text)
# Cosine similarity
logits_per_image, logits_per_text = model(image, text)
probs = logits_per_image.softmax(dim=-1).cpu().numpy()
# Print results
labels = ["a dog", "a cat", "a bird", "a car"]
for label, prob in zip(labels, probs[0]):
print(f"{label}: {prob:.2%}")
```
## Available models
```python
# Models (sorted by size)
models = [
"RN50", # ResNet-50
"RN101", # ResNet-101
"ViT-B/32", # Vision Transformer (recommended)
"ViT-B/16", # Better quality, slower
"ViT-L/14", # Best quality, slowest
]
model, preprocess = clip.load("ViT-B/32")
```
| Model | Parameters | Speed | Quality |
|-------|------------|-------|---------|
| RN50 | 102M | Fast | Good |
| ViT-B/32 | 151M | Medium | Better |
| ViT-L/14 | 428M | Slow | Best |
## Image-text similarity
```python
# Compute embeddings
image_features = model.encode_image(image)
text_features = model.encode_text(text)
# Normalize
image_features /= image_features.norm(dim=-1, keepdim=True)
text_features /= text_features.norm(dim=-1, keepdim=True)
# Cosine similarity
similarity = (image_features @ text_features.T).item()
print(f"Similarity: {similarity:.4f}")
```
## Semantic image search
```python
# Index images
image_paths = ["img1.jpg", "img2.jpg", "img3.jpg"]
image_embeddings = []
for img_path in image_paths:
image = preprocess(Image.open(img_path)).unsqueeze(0).to(device)
with torch.no_grad():
embedding = model.encode_image(image)
embedding /= embedding.norm(dim=-1, keepdim=True)
image_embeddings.append(embedding)
image_embeddings = torch.cat(image_embeddings)
# Search with text query
query = "a sunset over the ocean"
text_input = clip.tokenize([query]).to(device)
with torch.no_grad():
text_embedding = model.encode_text(text_input)
text_embedding /= text_embedding.norm(dim=-1, keepdim=True)
# Find most similar images
similarities = (text_embedding @ image_embeddings.T).squeeze(0)
top_k = similarities.topk(3)
for idx, score in zip(top_k.indices, top_k.values):
print(f"{image_paths[idx]}: {score:.3f}")
```
## Content moderation
```python
# Define categories
categories = [
"safe for work",
"not safe for work",
"violent content",
"graphic content"
]
text = clip.tokenize(categories).to(device)
# Check image
with torch.no_grad():
logits_per_image, _ = model(image, text)
probs = logits_per_image.softmax(dim=-1)
# Get classification
max_idx = probs.argmax().item()
max_prob = probs[0, max_idx].item()
print(f"Category: {categories[max_idx]} ({max_prob:.2%})")
```
## Batch processing
```python
# Process multiple images
images = [preprocess(Image.open(f"img{i}.jpg")) for i in range(10)]
images = torch.stack(images).to(device)
with torch.no_grad():
image_features = model.encode_image(images)
image_features /= image_features.norm(dim=-1, keepdim=True)
# Batch text
texts = ["a dog", "a cat", "a bird"]
text_tokens = clip.tokenize(texts).to(device)
with torch.no_grad():
text_features = model.encode_text(text_tokens)
text_features /= text_features.norm(dim=-1, keepdim=True)
# Similarity matrix (10 images × 3 texts)
similarities = image_features @ text_features.T
print(similarities.shape) # (10, 3)
```
## Integration with vector databases
```python
# Store CLIP embeddings in Chroma/FAISS
import chromadb
client = chromadb.Client()
collection = client.create_collection("image_embeddings")
# Add image embeddings
for img_path, embedding in zip(image_paths, image_embeddings):
collection.add(
embeddings=[embedding.cpu().numpy().tolist()],
metadatas=[{"path": img_path}],
ids=[img_path]
)
# Query with text
query = "a sunset"
text_embedding = model.encode_text(clip.tokenize([query]))
results = collection.query(
query_embeddings=[text_embedding.cpu().numpy().tolist()],
n_results=5
)
```
## Best practices
1. **Use ViT-B/32 for most cases** - Good balance
2. **Normalize embeddings** - Required for cosine similarity
3. **Batch processing** - More efficient
4. **Cache embeddings** - Expensive to recompute
5. **Use descriptive labels** - Better zero-shot performance
6. **GPU recommended** - 10-50× faster
7. **Preprocess images** - Use provided preprocess function
## Performance
| Operation | CPU | GPU (V100) |
|-----------|-----|------------|
| Image encoding | ~200ms | ~20ms |
| Text encoding | ~50ms | ~5ms |
| Similarity compute | &lt;1ms | &lt;1ms |
## Limitations
1. **Not for fine-grained tasks** - Best for broad categories
2. **Requires descriptive text** - Vague labels perform poorly
3. **Biased on web data** - May have dataset biases
4. **No bounding boxes** - Whole image only
5. **Limited spatial understanding** - Position/counting weak
## Resources
- **GitHub**: https://github.com/openai/CLIP ⭐ 25,300+
- **Paper**: https://arxiv.org/abs/2103.00020
- **Colab**: https://colab.research.google.com/github/openai/clip/
- **License**: MIT

239
website/docs/user-guide/skills/optional/mlops/mlops-faiss.md Normal file
View file

@ -0,0 +1,239 @@
---
title: "Faiss — Facebook's library for efficient similarity search and clustering of dense vectors"
sidebar_label: "Faiss"
description: "Facebook's library for efficient similarity search and clustering of dense vectors"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Faiss
Facebook's library for efficient similarity search and clustering of dense vectors. Supports billions of vectors, GPU acceleration, and various index types (Flat, IVF, HNSW). Use for fast k-NN search, large-scale vector retrieval, or when you need pure similarity search without metadata. Best for high-performance applications.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/faiss` |
| Path | `optional-skills/mlops/faiss` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `faiss-cpu`, `faiss-gpu`, `numpy` |
| Tags | `RAG`, `FAISS`, `Similarity Search`, `Vector Search`, `Facebook AI`, `GPU Acceleration`, `Billion-Scale`, `K-NN`, `HNSW`, `High Performance`, `Large Scale` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# FAISS - Efficient Similarity Search
Facebook AI's library for billion-scale vector similarity search.
## When to use FAISS
**Use FAISS when:**
- Need fast similarity search on large vector datasets (millions/billions)
- GPU acceleration required
- Pure vector similarity (no metadata filtering needed)
- High throughput, low latency critical
- Offline/batch processing of embeddings
**Metrics**:
- **31,700+ GitHub stars**
- Meta/Facebook AI Research
- **Handles billions of vectors**
- **C++** with Python bindings
**Use alternatives instead**:
- **Chroma/Pinecone**: Need metadata filtering
- **Weaviate**: Need full database features
- **Annoy**: Simpler, fewer features
## Quick start
### Installation
```bash
# CPU only
pip install faiss-cpu
# GPU support
pip install faiss-gpu
```
### Basic usage
```python
import faiss
import numpy as np
# Create sample data (1000 vectors, 128 dimensions)
d = 128
nb = 1000
vectors = np.random.random((nb, d)).astype('float32')
# Create index
index = faiss.IndexFlatL2(d) # L2 distance
index.add(vectors) # Add vectors
# Search
k = 5 # Find 5 nearest neighbors
query = np.random.random((1, d)).astype('float32')
distances, indices = index.search(query, k)
print(f"Nearest neighbors: {indices}")
print(f"Distances: {distances}")
```
## Index types
### 1. Flat (exact search)
```python
# L2 (Euclidean) distance
index = faiss.IndexFlatL2(d)
# Inner product (cosine similarity if normalized)
index = faiss.IndexFlatIP(d)
# Slowest, most accurate
```
### 2. IVF (inverted file) - Fast approximate
```python
# Create quantizer
quantizer = faiss.IndexFlatL2(d)
# IVF index with 100 clusters
nlist = 100
index = faiss.IndexIVFFlat(quantizer, d, nlist)
# Train on data
index.train(vectors)
# Add vectors
index.add(vectors)
# Search (nprobe = clusters to search)
index.nprobe = 10
distances, indices = index.search(query, k)
```
### 3. HNSW (Hierarchical NSW) - Best quality/speed
```python
# HNSW index
M = 32 # Number of connections per layer
index = faiss.IndexHNSWFlat(d, M)
# No training needed
index.add(vectors)
# Search
distances, indices = index.search(query, k)
```
### 4. Product Quantization - Memory efficient
```python
# PQ reduces memory by 16-32×
m = 8 # Number of subquantizers
nbits = 8
index = faiss.IndexPQ(d, m, nbits)
# Train and add
index.train(vectors)
index.add(vectors)
```
## Save and load
```python
# Save index
faiss.write_index(index, "large.index")
# Load index
index = faiss.read_index("large.index")
# Continue using
distances, indices = index.search(query, k)
```
## GPU acceleration
```python
# Single GPU
res = faiss.StandardGpuResources()
index_cpu = faiss.IndexFlatL2(d)
index_gpu = faiss.index_cpu_to_gpu(res, 0, index_cpu) # GPU 0
# Multi-GPU
index_gpu = faiss.index_cpu_to_all_gpus(index_cpu)
# 10-100× faster than CPU
```
## LangChain integration
```python
from langchain_community.vectorstores import FAISS
from langchain_openai import OpenAIEmbeddings
# Create FAISS vector store
vectorstore = FAISS.from_documents(docs, OpenAIEmbeddings())
# Save
vectorstore.save_local("faiss_index")
# Load
vectorstore = FAISS.load_local(
"faiss_index",
OpenAIEmbeddings(),
allow_dangerous_deserialization=True
)
# Search
results = vectorstore.similarity_search("query", k=5)
```
## LlamaIndex integration
```python
from llama_index.vector_stores.faiss import FaissVectorStore
import faiss
# Create FAISS index
d = 1536
faiss_index = faiss.IndexFlatL2(d)
vector_store = FaissVectorStore(faiss_index=faiss_index)
```
## Best practices
1. **Choose right index type** - Flat for &lt;10K, IVF for 10K-1M, HNSW for quality
2. **Normalize for cosine** - Use IndexFlatIP with normalized vectors
3. **Use GPU for large datasets** - 10-100× faster
4. **Save trained indices** - Training is expensive
5. **Tune nprobe/ef_search** - Balance speed/accuracy
6. **Monitor memory** - PQ for large datasets
7. **Batch queries** - Better GPU utilization
## Performance
| Index Type | Build Time | Search Time | Memory | Accuracy |
|------------|------------|-------------|--------|----------|
| Flat | Fast | Slow | High | 100% |
| IVF | Medium | Fast | Medium | 95-99% |
| HNSW | Slow | Fastest | High | 99% |
| PQ | Medium | Fast | Low | 90-95% |
## Resources
- **GitHub**: https://github.com/facebookresearch/faiss ⭐ 31,700+
- **Wiki**: https://github.com/facebookresearch/faiss/wiki
- **License**: MIT

384
website/docs/user-guide/skills/optional/mlops/mlops-flash-attention.md Normal file
View file

@ -0,0 +1,384 @@
---
title: "Optimizing Attention Flash"
sidebar_label: "Optimizing Attention Flash"
description: "Optimizes transformer attention with Flash Attention for 2-4x speedup and 10-20x memory reduction"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Optimizing Attention Flash
Optimizes transformer attention with Flash Attention for 2-4x speedup and 10-20x memory reduction. Use when training/running transformers with long sequences (>512 tokens), encountering GPU memory issues with attention, or need faster inference. Supports PyTorch native SDPA, flash-attn library, H100 FP8, and sliding window attention.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/flash-attention` |
| Path | `optional-skills/mlops/flash-attention` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `flash-attn`, `torch`, `transformers` |
| Tags | `Optimization`, `Flash Attention`, `Attention Optimization`, `Memory Efficiency`, `Speed Optimization`, `Long Context`, `PyTorch`, `SDPA`, `H100`, `FP8`, `Transformers` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Flash Attention - Fast Memory-Efficient Attention
## Quick start
Flash Attention provides 2-4x speedup and 10-20x memory reduction for transformer attention through IO-aware tiling and recomputation.
**PyTorch native (easiest, PyTorch 2.2+)**:
```python
import torch
import torch.nn.functional as F
q = torch.randn(2, 8, 512, 64, device='cuda', dtype=torch.float16) # [batch, heads, seq, dim]
k = torch.randn(2, 8, 512, 64, device='cuda', dtype=torch.float16)
v = torch.randn(2, 8, 512, 64, device='cuda', dtype=torch.float16)
# Automatically uses Flash Attention if available
out = F.scaled_dot_product_attention(q, k, v)
```
**flash-attn library (more features)**:
```bash
pip install flash-attn --no-build-isolation
```
```python
from flash_attn import flash_attn_func
# q, k, v: [batch, seqlen, nheads, headdim]
out = flash_attn_func(q, k, v, dropout_p=0.0, causal=True)
```
## Common workflows
### Workflow 1: Enable in existing PyTorch model
Copy this checklist:
```
Flash Attention Integration:
- [ ] Step 1: Check PyTorch version (≥2.2)
- [ ] Step 2: Enable Flash Attention backend
- [ ] Step 3: Verify speedup with profiling
- [ ] Step 4: Test accuracy matches baseline
```
**Step 1: Check PyTorch version**
```bash
python -c "import torch; print(torch.__version__)"
# Should be ≥2.2.0
```
If &lt;2.2, upgrade:
```bash
pip install --upgrade torch
```
**Step 2: Enable Flash Attention backend**
Replace standard attention:
```python
# Before (standard attention)
attn_weights = torch.softmax(q @ k.transpose(-2, -1) / math.sqrt(d_k), dim=-1)
out = attn_weights @ v
# After (Flash Attention)
import torch.nn.functional as F
out = F.scaled_dot_product_attention(q, k, v, attn_mask=mask)
```
Force Flash Attention backend:
```python
with torch.backends.cuda.sdp_kernel(
enable_flash=True,
enable_math=False,
enable_mem_efficient=False
):
out = F.scaled_dot_product_attention(q, k, v)
```
**Step 3: Verify speedup with profiling**
```python
import torch.utils.benchmark as benchmark
def test_attention(use_flash):
q, k, v = [torch.randn(2, 8, 2048, 64, device='cuda', dtype=torch.float16) for _ in range(3)]
if use_flash:
with torch.backends.cuda.sdp_kernel(enable_flash=True):
return F.scaled_dot_product_attention(q, k, v)
else:
attn = (q @ k.transpose(-2, -1) / 8.0).softmax(dim=-1)
return attn @ v
# Benchmark
t_flash = benchmark.Timer(stmt='test_attention(True)', globals=globals())
t_standard = benchmark.Timer(stmt='test_attention(False)', globals=globals())
print(f"Flash: {t_flash.timeit(100).mean:.3f}s")
print(f"Standard: {t_standard.timeit(100).mean:.3f}s")
```
Expected: 2-4x speedup for sequences >512 tokens.
**Step 4: Test accuracy matches baseline**
```python
# Compare outputs
q, k, v = [torch.randn(1, 8, 512, 64, device='cuda', dtype=torch.float16) for _ in range(3)]
# Flash Attention
out_flash = F.scaled_dot_product_attention(q, k, v)
# Standard attention
attn_weights = torch.softmax(q @ k.transpose(-2, -1) / 8.0, dim=-1)
out_standard = attn_weights @ v
# Check difference
diff = (out_flash - out_standard).abs().max()
print(f"Max difference: {diff:.6f}")
# Should be <1e-3 for float16
```
### Workflow 2: Use flash-attn library for advanced features
For multi-query attention, sliding window, or H100 FP8.
Copy this checklist:
```
flash-attn Library Setup:
- [ ] Step 1: Install flash-attn library
- [ ] Step 2: Modify attention code
- [ ] Step 3: Enable advanced features
- [ ] Step 4: Benchmark performance
```
**Step 1: Install flash-attn library**
```bash
# NVIDIA GPUs (CUDA 12.0+)
pip install flash-attn --no-build-isolation
# Verify installation
python -c "from flash_attn import flash_attn_func; print('Success')"
```
**Step 2: Modify attention code**
```python
from flash_attn import flash_attn_func
# Input: [batch_size, seq_len, num_heads, head_dim]
# Transpose from [batch, heads, seq, dim] if needed
q = q.transpose(1, 2) # [batch, seq, heads, dim]
k = k.transpose(1, 2)
v = v.transpose(1, 2)
out = flash_attn_func(
q, k, v,
dropout_p=0.1,
causal=True, # For autoregressive models
window_size=(-1, -1), # No sliding window
softmax_scale=None # Auto-scale
)
out = out.transpose(1, 2) # Back to [batch, heads, seq, dim]
```
**Step 3: Enable advanced features**
Multi-query attention (shared K/V across heads):
```python
from flash_attn import flash_attn_func
# q: [batch, seq, num_q_heads, dim]
# k, v: [batch, seq, num_kv_heads, dim] # Fewer KV heads
out = flash_attn_func(q, k, v) # Automatically handles MQA
```
Sliding window attention (local attention):
```python
# Only attend to window of 256 tokens before/after
out = flash_attn_func(
q, k, v,
window_size=(256, 256), # (left, right) window
causal=True
)
```
**Step 4: Benchmark performance**
```python
import torch
from flash_attn import flash_attn_func
import time
q, k, v = [torch.randn(4, 4096, 32, 64, device='cuda', dtype=torch.float16) for _ in range(3)]
# Warmup
for _ in range(10):
_ = flash_attn_func(q, k, v)
# Benchmark
torch.cuda.synchronize()
start = time.time()
for _ in range(100):
out = flash_attn_func(q, k, v)
torch.cuda.synchronize()
end = time.time()
print(f"Time per iteration: {(end-start)/100*1000:.2f}ms")
print(f"Memory allocated: {torch.cuda.max_memory_allocated()/1e9:.2f}GB")
```
### Workflow 3: H100 FP8 optimization (FlashAttention-3)
For maximum performance on H100 GPUs.
```
FP8 Setup:
- [ ] Step 1: Verify H100 GPU available
- [ ] Step 2: Install flash-attn with FP8 support
- [ ] Step 3: Convert inputs to FP8
- [ ] Step 4: Run with FP8 attention
```
**Step 1: Verify H100 GPU**
```bash
nvidia-smi --query-gpu=name --format=csv
# Should show "H100" or "H800"
```
**Step 2: Install flash-attn with FP8 support**
```bash
pip install flash-attn --no-build-isolation
# FP8 support included for H100
```
**Step 3: Convert inputs to FP8**
```python
import torch
q = torch.randn(2, 4096, 32, 64, device='cuda', dtype=torch.float16)
k = torch.randn(2, 4096, 32, 64, device='cuda', dtype=torch.float16)
v = torch.randn(2, 4096, 32, 64, device='cuda', dtype=torch.float16)
# Convert to float8_e4m3 (FP8)
q_fp8 = q.to(torch.float8_e4m3fn)
k_fp8 = k.to(torch.float8_e4m3fn)
v_fp8 = v.to(torch.float8_e4m3fn)
```
**Step 4: Run with FP8 attention**
```python
from flash_attn import flash_attn_func
# FlashAttention-3 automatically uses FP8 kernels on H100
out = flash_attn_func(q_fp8, k_fp8, v_fp8)
# Result: ~1.2 PFLOPS, 1.5-2x faster than FP16
```
## When to use vs alternatives
**Use Flash Attention when:**
- Training transformers with sequences >512 tokens
- Running inference with long context (>2K tokens)
- GPU memory constrained (OOM with standard attention)
- Need 2-4x speedup without accuracy loss
- Using PyTorch 2.2+ or can install flash-attn
**Use alternatives instead:**
- **Standard attention**: Sequences &lt;256 tokens (overhead not worth it)
- **xFormers**: Need more attention variants (not just speed)
- **Memory-efficient attention**: CPU inference (Flash Attention needs GPU)
## Common issues
**Issue: ImportError: cannot import flash_attn**
Install with no-build-isolation flag:
```bash
pip install flash-attn --no-build-isolation
```
Or install CUDA toolkit first:
```bash
conda install cuda -c nvidia
pip install flash-attn --no-build-isolation
```
**Issue: Slower than expected (no speedup)**
Flash Attention benefits increase with sequence length:
- &lt;512 tokens: Minimal speedup (10-20%)
- 512-2K tokens: 2-3x speedup
- >2K tokens: 3-4x speedup
Check sequence length is sufficient.
**Issue: RuntimeError: CUDA error**
Verify GPU supports Flash Attention:
```python
import torch
print(torch.cuda.get_device_capability())
# Should be ≥(7, 5) for Turing+
```
Flash Attention requires:
- Ampere (A100, A10): ✅ Full support
- Turing (T4): ✅ Supported
- Volta (V100): ❌ Not supported
**Issue: Accuracy degradation**
Check dtype is float16 or bfloat16 (not float32):
```python
q = q.to(torch.float16) # Or torch.bfloat16
```
Flash Attention uses float16/bfloat16 for speed. Float32 not supported.
## Advanced topics
**Integration with HuggingFace Transformers**: See [references/transformers-integration.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/flash-attention/references/transformers-integration.md) for enabling Flash Attention in BERT, GPT, Llama models.
**Performance benchmarks**: See [references/benchmarks.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/flash-attention/references/benchmarks.md) for detailed speed and memory comparisons across GPUs and sequence lengths.
**Algorithm details**: See [references/algorithm.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/flash-attention/references/algorithm.md) for tiling strategy, recomputation, and IO complexity analysis.
**Advanced features**: See [references/advanced-features.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/flash-attention/references/advanced-features.md) for rotary embeddings, ALiBi, paged KV cache, and custom attention masks.
## Hardware requirements
- **GPU**: NVIDIA Ampere+ (A100, A10, A30) or AMD MI200+
- **VRAM**: Same as standard attention (Flash Attention doesn't increase memory)
- **CUDA**: 12.0+ (11.8 minimum)
- **PyTorch**: 2.2+ for native support
**Not supported**: V100 (Volta), CPU inference
## Resources
- Paper: "FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness" (NeurIPS 2022)
- Paper: "FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning" (ICLR 2024)
- Blog: https://tridao.me/blog/2024/flash3/
- GitHub: https://github.com/Dao-AILab/flash-attention
- PyTorch docs: https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html

590
website/docs/user-guide/skills/optional/mlops/mlops-guidance.md Normal file
View file

@ -0,0 +1,590 @@
---
title: "Guidance"
sidebar_label: "Guidance"
description: "Control LLM output with regex and grammars, guarantee valid JSON/XML/code generation, enforce structured formats, and build multi-step workflows with Guidanc..."
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Guidance
Control LLM output with regex and grammars, guarantee valid JSON/XML/code generation, enforce structured formats, and build multi-step workflows with Guidance - Microsoft Research's constrained generation framework
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/guidance` |
| Path | `optional-skills/mlops/guidance` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `guidance`, `transformers` |
| Tags | `Prompt Engineering`, `Guidance`, `Constrained Generation`, `Structured Output`, `JSON Validation`, `Grammar`, `Microsoft Research`, `Format Enforcement`, `Multi-Step Workflows` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Guidance: Constrained LLM Generation
## When to Use This Skill
Use Guidance when you need to:
- **Control LLM output syntax** with regex or grammars
- **Guarantee valid JSON/XML/code** generation
- **Reduce latency** vs traditional prompting approaches
- **Enforce structured formats** (dates, emails, IDs, etc.)
- **Build multi-step workflows** with Pythonic control flow
- **Prevent invalid outputs** through grammatical constraints
**GitHub Stars**: 18,000+ | **From**: Microsoft Research
## Installation
```bash
# Base installation
pip install guidance
# With specific backends
pip install guidance[transformers] # Hugging Face models
pip install guidance[llama_cpp] # llama.cpp models
```
## Quick Start
### Basic Example: Structured Generation
```python
from guidance import models, gen
# Load model (supports OpenAI, Transformers, llama.cpp)
lm = models.OpenAI("gpt-4")
# Generate with constraints
result = lm + "The capital of France is " + gen("capital", max_tokens=5)
print(result["capital"]) # "Paris"
```
### With Anthropic Claude
```python
from guidance import models, gen, system, user, assistant
# Configure Claude
lm = models.Anthropic("claude-sonnet-4-5-20250929")
# Use context managers for chat format
with system():
lm += "You are a helpful assistant."
with user():
lm += "What is the capital of France?"
with assistant():
lm += gen(max_tokens=20)
```
## Core Concepts
### 1. Context Managers
Guidance uses Pythonic context managers for chat-style interactions.
```python
from guidance import system, user, assistant, gen
lm = models.Anthropic("claude-sonnet-4-5-20250929")
# System message
with system():
lm += "You are a JSON generation expert."
# User message
with user():
lm += "Generate a person object with name and age."
# Assistant response
with assistant():
lm += gen("response", max_tokens=100)
print(lm["response"])
```
**Benefits:**
- Natural chat flow
- Clear role separation
- Easy to read and maintain
### 2. Constrained Generation
Guidance ensures outputs match specified patterns using regex or grammars.
#### Regex Constraints
```python
from guidance import models, gen
lm = models.Anthropic("claude-sonnet-4-5-20250929")
# Constrain to valid email format
lm += "Email: " + gen("email", regex=r"[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}")
# Constrain to date format (YYYY-MM-DD)
lm += "Date: " + gen("date", regex=r"\d{4}-\d{2}-\d{2}")
# Constrain to phone number
lm += "Phone: " + gen("phone", regex=r"\d{3}-\d{3}-\d{4}")
print(lm["email"]) # Guaranteed valid email
print(lm["date"]) # Guaranteed YYYY-MM-DD format
```
**How it works:**
- Regex converted to grammar at token level
- Invalid tokens filtered during generation
- Model can only produce matching outputs
#### Selection Constraints
```python
from guidance import models, gen, select
lm = models.Anthropic("claude-sonnet-4-5-20250929")
# Constrain to specific choices
lm += "Sentiment: " + select(["positive", "negative", "neutral"], name="sentiment")
# Multiple-choice selection
lm += "Best answer: " + select(
["A) Paris", "B) London", "C) Berlin", "D) Madrid"],
name="answer"
)
print(lm["sentiment"]) # One of: positive, negative, neutral
print(lm["answer"]) # One of: A, B, C, or D
```
### 3. Token Healing
Guidance automatically "heals" token boundaries between prompt and generation.
**Problem:** Tokenization creates unnatural boundaries.
```python
# Without token healing
prompt = "The capital of France is "
# Last token: " is "
# First generated token might be " Par" (with leading space)
# Result: "The capital of France is Paris" (double space!)
```
**Solution:** Guidance backs up one token and regenerates.
```python
from guidance import models, gen
lm = models.Anthropic("claude-sonnet-4-5-20250929")
# Token healing enabled by default
lm += "The capital of France is " + gen("capital", max_tokens=5)
# Result: "The capital of France is Paris" (correct spacing)
```
**Benefits:**
- Natural text boundaries
- No awkward spacing issues
- Better model performance (sees natural token sequences)
### 4. Grammar-Based Generation
Define complex structures using context-free grammars.
```python
from guidance import models, gen
lm = models.Anthropic("claude-sonnet-4-5-20250929")
# JSON grammar (simplified)
json_grammar = """
{
"name": <gen name regex="[A-Za-z ]+" max_tokens=20>,
"age": <gen age regex="[0-9]+" max_tokens=3>,
"email": <gen email regex="[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\\.[a-zA-Z]{2,}" max_tokens=50>
}
"""
# Generate valid JSON
lm += gen("person", grammar=json_grammar)
print(lm["person"]) # Guaranteed valid JSON structure
```
**Use cases:**
- Complex structured outputs
- Nested data structures
- Programming language syntax
- Domain-specific languages
### 5. Guidance Functions
Create reusable generation patterns with the `@guidance` decorator.
```python
from guidance import guidance, gen, models
@guidance
def generate_person(lm):
"""Generate a person with name and age."""
lm += "Name: " + gen("name", max_tokens=20, stop="\n")
lm += "\nAge: " + gen("age", regex=r"[0-9]+", max_tokens=3)
return lm
# Use the function
lm = models.Anthropic("claude-sonnet-4-5-20250929")
lm = generate_person(lm)
print(lm["name"])
print(lm["age"])
```
**Stateful Functions:**
```python
@guidance(stateless=False)
def react_agent(lm, question, tools, max_rounds=5):
"""ReAct agent with tool use."""
lm += f"Question: {question}\n\n"
for i in range(max_rounds):
# Thought
lm += f"Thought {i+1}: " + gen("thought", stop="\n")
# Action
lm += "\nAction: " + select(list(tools.keys()), name="action")
# Execute tool
tool_result = tools[lm["action"]]()
lm += f"\nObservation: {tool_result}\n\n"
# Check if done
lm += "Done? " + select(["Yes", "No"], name="done")
if lm["done"] == "Yes":
break
# Final answer
lm += "\nFinal Answer: " + gen("answer", max_tokens=100)
return lm
```
## Backend Configuration
### Anthropic Claude
```python
from guidance import models
lm = models.Anthropic(
model="claude-sonnet-4-5-20250929",
api_key="your-api-key" # Or set ANTHROPIC_API_KEY env var
)
```
### OpenAI
```python
lm = models.OpenAI(
model="gpt-4o-mini",
api_key="your-api-key" # Or set OPENAI_API_KEY env var
)
```
### Local Models (Transformers)
```python
from guidance.models import Transformers
lm = Transformers(
"microsoft/Phi-4-mini-instruct",
device="cuda" # Or "cpu"
)
```
### Local Models (llama.cpp)
```python
from guidance.models import LlamaCpp
lm = LlamaCpp(
model_path="/path/to/model.gguf",
n_ctx=4096,
n_gpu_layers=35
)
```
## Common Patterns
### Pattern 1: JSON Generation
```python
from guidance import models, gen, system, user, assistant
lm = models.Anthropic("claude-sonnet-4-5-20250929")
with system():
lm += "You generate valid JSON."
with user():
lm += "Generate a user profile with name, age, and email."
with assistant():
lm += """{
"name": """ + gen("name", regex=r'"[A-Za-z ]+"', max_tokens=30) + """,
"age": """ + gen("age", regex=r"[0-9]+", max_tokens=3) + """,
"email": """ + gen("email", regex=r'"[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}"', max_tokens=50) + """
}"""
print(lm) # Valid JSON guaranteed
```
### Pattern 2: Classification
```python
from guidance import models, gen, select
lm = models.Anthropic("claude-sonnet-4-5-20250929")
text = "This product is amazing! I love it."
lm += f"Text: {text}\n"
lm += "Sentiment: " + select(["positive", "negative", "neutral"], name="sentiment")
lm += "\nConfidence: " + gen("confidence", regex=r"[0-9]+", max_tokens=3) + "%"
print(f"Sentiment: {lm['sentiment']}")
print(f"Confidence: {lm['confidence']}%")
```
### Pattern 3: Multi-Step Reasoning
```python
from guidance import models, gen, guidance
@guidance
def chain_of_thought(lm, question):
"""Generate answer with step-by-step reasoning."""
lm += f"Question: {question}\n\n"
# Generate multiple reasoning steps
for i in range(3):
lm += f"Step {i+1}: " + gen(f"step_{i+1}", stop="\n", max_tokens=100) + "\n"
# Final answer
lm += "\nTherefore, the answer is: " + gen("answer", max_tokens=50)
return lm
lm = models.Anthropic("claude-sonnet-4-5-20250929")
lm = chain_of_thought(lm, "What is 15% of 200?")
print(lm["answer"])
```
### Pattern 4: ReAct Agent
```python
from guidance import models, gen, select, guidance
@guidance(stateless=False)
def react_agent(lm, question):
"""ReAct agent with tool use."""
tools = {
"calculator": lambda expr: eval(expr),
"search": lambda query: f"Search results for: {query}",
}
lm += f"Question: {question}\n\n"
for round in range(5):
# Thought
lm += f"Thought: " + gen("thought", stop="\n") + "\n"
# Action selection
lm += "Action: " + select(["calculator", "search", "answer"], name="action")
if lm["action"] == "answer":
lm += "\nFinal Answer: " + gen("answer", max_tokens=100)
break
# Action input
lm += "\nAction Input: " + gen("action_input", stop="\n") + "\n"
# Execute tool
if lm["action"] in tools:
result = tools[lm["action"]](lm["action_input"])
lm += f"Observation: {result}\n\n"
return lm
lm = models.Anthropic("claude-sonnet-4-5-20250929")
lm = react_agent(lm, "What is 25 * 4 + 10?")
print(lm["answer"])
```
### Pattern 5: Data Extraction
```python
from guidance import models, gen, guidance
@guidance
def extract_entities(lm, text):
"""Extract structured entities from text."""
lm += f"Text: {text}\n\n"
# Extract person
lm += "Person: " + gen("person", stop="\n", max_tokens=30) + "\n"
# Extract organization
lm += "Organization: " + gen("organization", stop="\n", max_tokens=30) + "\n"
# Extract date
lm += "Date: " + gen("date", regex=r"\d{4}-\d{2}-\d{2}", max_tokens=10) + "\n"
# Extract location
lm += "Location: " + gen("location", stop="\n", max_tokens=30) + "\n"
return lm
text = "Tim Cook announced at Apple Park on 2024-09-15 in Cupertino."
lm = models.Anthropic("claude-sonnet-4-5-20250929")
lm = extract_entities(lm, text)
print(f"Person: {lm['person']}")
print(f"Organization: {lm['organization']}")
print(f"Date: {lm['date']}")
print(f"Location: {lm['location']}")
```
## Best Practices
### 1. Use Regex for Format Validation
```python
# ✅ Good: Regex ensures valid format
lm += "Email: " + gen("email", regex=r"[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}")
# ❌ Bad: Free generation may produce invalid emails
lm += "Email: " + gen("email", max_tokens=50)
```
### 2. Use select() for Fixed Categories
```python
# ✅ Good: Guaranteed valid category
lm += "Status: " + select(["pending", "approved", "rejected"], name="status")
# ❌ Bad: May generate typos or invalid values
lm += "Status: " + gen("status", max_tokens=20)
```
### 3. Leverage Token Healing
```python
# Token healing is enabled by default
# No special action needed - just concatenate naturally
lm += "The capital is " + gen("capital") # Automatic healing
```
### 4. Use stop Sequences
```python
# ✅ Good: Stop at newline for single-line outputs
lm += "Name: " + gen("name", stop="\n")
# ❌ Bad: May generate multiple lines
lm += "Name: " + gen("name", max_tokens=50)
```
### 5. Create Reusable Functions
```python
# ✅ Good: Reusable pattern
@guidance
def generate_person(lm):
lm += "Name: " + gen("name", stop="\n")
lm += "\nAge: " + gen("age", regex=r"[0-9]+")
return lm
# Use multiple times
lm = generate_person(lm)
lm += "\n\n"
lm = generate_person(lm)
```
### 6. Balance Constraints
```python
# ✅ Good: Reasonable constraints
lm += gen("name", regex=r"[A-Za-z ]+", max_tokens=30)
# ❌ Too strict: May fail or be very slow
lm += gen("name", regex=r"^(John|Jane)$", max_tokens=10)
```
## Comparison to Alternatives
| Feature | Guidance | Instructor | Outlines | LMQL |
|---------|----------|------------|----------|------|
| Regex Constraints | ✅ Yes | ❌ No | ✅ Yes | ✅ Yes |
| Grammar Support | ✅ CFG | ❌ No | ✅ CFG | ✅ CFG |
| Pydantic Validation | ❌ No | ✅ Yes | ✅ Yes | ❌ No |
| Token Healing | ✅ Yes | ❌ No | ✅ Yes | ❌ No |
| Local Models | ✅ Yes | ⚠️ Limited | ✅ Yes | ✅ Yes |
| API Models | ✅ Yes | ✅ Yes | ⚠️ Limited | ✅ Yes |
| Pythonic Syntax | ✅ Yes | ✅ Yes | ✅ Yes | ❌ SQL-like |
| Learning Curve | Low | Low | Medium | High |
**When to choose Guidance:**
- Need regex/grammar constraints
- Want token healing
- Building complex workflows with control flow
- Using local models (Transformers, llama.cpp)
- Prefer Pythonic syntax
**When to choose alternatives:**
- Instructor: Need Pydantic validation with automatic retrying
- Outlines: Need JSON schema validation
- LMQL: Prefer declarative query syntax
## Performance Characteristics
**Latency Reduction:**
- 30-50% faster than traditional prompting for constrained outputs
- Token healing reduces unnecessary regeneration
- Grammar constraints prevent invalid token generation
**Memory Usage:**
- Minimal overhead vs unconstrained generation
- Grammar compilation cached after first use
- Efficient token filtering at inference time
**Token Efficiency:**
- Prevents wasted tokens on invalid outputs
- No need for retry loops
- Direct path to valid outputs
## Resources
- **Documentation**: https://guidance.readthedocs.io
- **GitHub**: https://github.com/guidance-ai/guidance (18k+ stars)
- **Notebooks**: https://github.com/guidance-ai/guidance/tree/main/notebooks
- **Discord**: Community support available
## See Also
- `references/constraints.md` - Comprehensive regex and grammar patterns
- `references/backends.md` - Backend-specific configuration
- `references/examples.md` - Production-ready examples

320
website/docs/user-guide/skills/optional/mlops/mlops-hermes-atropos-environments.md Normal file
View file

@ -0,0 +1,320 @@
---
title: "Hermes Atropos Environments — Build, test, and debug Hermes Agent RL environments for Atropos training"
sidebar_label: "Hermes Atropos Environments"
description: "Build, test, and debug Hermes Agent RL environments for Atropos training"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Hermes Atropos Environments
Build, test, and debug Hermes Agent RL environments for Atropos training. Covers the HermesAgentBaseEnv interface, reward functions, agent loop integration, evaluation with tools, wandb logging, and the three CLI modes (serve/process/evaluate). Use when creating, reviewing, or fixing RL environments in the hermes-agent repo.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/hermes-atropos-environments` |
| Path | `optional-skills/mlops/hermes-atropos-environments` |
| Version | `1.1.0` |
| Author | Hermes Agent |
| License | MIT |
| Tags | `atropos`, `rl`, `environments`, `training`, `reinforcement-learning`, `reward-functions` |
| Related skills | [`axolotl`](/docs/user-guide/skills/bundled/mlops/mlops-training-axolotl), [`fine-tuning-with-trl`](/docs/user-guide/skills/bundled/mlops/mlops-training-trl-fine-tuning), `lm-evaluation-harness` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Hermes Agent Atropos Environments
Guide for building RL environments in the hermes-agent repo that integrate with the Atropos training framework.
## Architecture Overview
```
Atropos BaseEnv (atroposlib/envs/base.py)
└── HermesAgentBaseEnv (environments/hermes_base_env.py)
├── Handles agent loop orchestration
├── Handles tool resolution per group
├── Handles ToolContext for reward verification
└── YOUR ENVIRONMENT (environments/your_env.py)
Only implements: setup, get_next_item, format_prompt,
compute_reward, evaluate, wandb_log
```
Hermes environments are special because they run a **multi-turn agent loop with tool calling** — not just single-turn completions. The base env handles the loop; you implement the task and scoring.
## File Locations
| File | Purpose |
|------|---------|
| `environments/hermes_base_env.py` | Base class with agent loop + tool resolution |
| `environments/agent_loop.py` | `HermesAgentLoop` + `AgentResult` dataclass |
| `environments/tool_context.py` | `ToolContext` for reward verification |
| `environments/tool_call_parsers.py` | Phase 2 tool call parsers (hermes, mistral, etc.) |
| `environments/your_env.py` | Your environment implementation |
## Inference Setup — Ask the User First
**IMPORTANT:** Before running any test, evaluation, or data generation command, always ask the user how they want to handle inference. Do NOT assume OpenRouter or any specific endpoint. Present these options:
1. **OpenRouter** — Ask which model they want to use (e.g., `anthropic/claude-sonnet-4.5`, `google/gemini-2.5-pro`, `meta-llama/llama-3.3-70b-instruct`, etc.). Requires `OPENROUTER_API_KEY` in environment.
2. **Self-hosted VLLM endpoint** — Ask for their base URL (e.g., `http://localhost:8000/v1`) and model name. Set `--openai.server_type vllm`.
3. **Other OpenAI-compatible API** — Ask for the base URL, model name, and any required API key. Set `--openai.server_type openai` and `--openai.health_check false`.
4. **Local Atropos training server** — For `serve` mode with a live training loop. Default `http://localhost:8000/v1`.
Once the user tells you their setup, use those values in all CLI commands for that session. Example prompts:
> "Before I run this, how would you like to handle inference?
> 1. OpenRouter (I'll need your preferred model, e.g. claude-sonnet-4.5)
> 2. A self-hosted VLLM endpoint (give me the URL and model name)
> 3. Another OpenAI-compatible API (give me the URL, model, and any auth details)
> 4. Local Atropos training server (serve mode)"
### Key flags by provider:
| Provider | `--openai.server_type` | `--openai.health_check` | `--openai.api_key` |
|----------|----------------------|------------------------|-------------------|
| OpenRouter | `openai` | `false` | `$OPENROUTER_API_KEY` |
| VLLM (self-hosted) | `vllm` | (default) | (not needed) |
| Other OpenAI-compatible | `openai` | `false` | As needed |
| Local Atropos | (default) | (default) | (not needed) |
## Required Methods
### 1. `setup()` — Load dataset and initialize state
```python
async def setup(self) -> None:
"""Called once at startup. Load datasets, initialize state."""
# Try HuggingFace first, fallback to built-in samples
try:
from datasets import load_dataset
ds = load_dataset("your/dataset", split="test")
self._items = [...]
except Exception:
self._items = BUILTIN_SAMPLES
# Always split into train/eval
random.shuffle(self._items)
eval_size = max(20, int(len(self._items) * 0.1))
self._eval_items = self._items[:eval_size]
self._items = self._items[eval_size:]
```
### 2. `get_next_item()` — Return next training item
```python
async def get_next_item(self) -> dict:
"""Return next item, cycling through dataset."""
item = self._items[self._index % len(self._items)]
self._index += 1
return item
```
### 3. `format_prompt(item)` — Convert item to user message
```python
def format_prompt(self, item: dict) -> str:
"""Convert a dataset item into the user-facing prompt."""
return f"Research this question: {item['question']}"
```
### 4. `compute_reward(item, result, ctx)` — Score the rollout
**CRITICAL**: `result` is an `AgentResult`, NOT a dict. It has these attributes:
- `result.messages` — List of message dicts (OpenAI format)
- `result.turns_used` — Number of LLM calls made
- `result.finished_naturally` — True if model stopped voluntarily
- `result.tool_errors` — List of ToolError objects
**AgentResult does NOT have**: `final_response`, `tool_calls`, `tools_used`.
You must extract these from `result.messages`:
```python
async def compute_reward(self, item, result: AgentResult, ctx: ToolContext) -> float:
# Extract final response (last assistant message with content)
final_response = ""
tools_used = []
for msg in reversed(result.messages):
if msg.get("role") == "assistant" and msg.get("content") and not final_response:
final_response = msg["content"]
if msg.get("role") == "assistant" and msg.get("tool_calls"):
for tc in msg["tool_calls"]:
fn = tc.get("function", {}) if isinstance(tc, dict) else {}
name = fn.get("name", "")
if name:
tools_used.append(name)
# Score using LLM judge, heuristic, or ToolContext verification
correctness = await self._llm_judge(item, final_response)
return correctness
```
`ctx` (ToolContext) gives you terminal/file access to the agent's sandbox for verification:
```python
# Run tests in the agent's sandbox
result = ctx.terminal("pytest /workspace/test.py")
return 1.0 if result["exit_code"] == 0 else 0.0
```
### 5. `evaluate()` — Periodic evaluation with full agent loop
**MUST use the full agent loop with tools**, not single-turn chat_completion.
The whole point of hermes-agent environments is agentic evaluation:
```python
async def evaluate(self, *args, **kwargs) -> None:
import time, uuid
from environments.agent_loop import HermesAgentLoop
from environments.tool_context import ToolContext
start_time = time.time()
tools, valid_names = self._resolve_tools_for_group()
samples = []
for item in self._eval_items[:self.config.eval_size]:
task_id = str(uuid.uuid4())
messages = []
if self.config.system_prompt:
messages.append({"role": "system", "content": self.config.system_prompt})
messages.append({"role": "user", "content": self.format_prompt(item)})
agent = HermesAgentLoop(
server=self.server,
tool_schemas=tools,
valid_tool_names=valid_names,
max_turns=self.config.max_agent_turns,
task_id=task_id,
temperature=0.0, # Deterministic for eval
max_tokens=self.config.max_token_length,
extra_body=self.config.extra_body,
)
result = await agent.run(messages)
ctx = ToolContext(task_id)
try:
reward = await self.compute_reward(item, result, ctx)
finally:
ctx.cleanup()
samples.append({"prompt": ..., "response": ..., "reward": reward})
eval_metrics = {"eval/mean_reward": ...}
await self.evaluate_log(metrics=eval_metrics, samples=samples,
start_time=start_time, end_time=time.time())
```
### 6. `wandb_log()` — Custom metrics logging
Always call `super().wandb_log()` at the end:
```python
async def wandb_log(self, wandb_metrics=None):
if wandb_metrics is None:
wandb_metrics = {}
if self._reward_buffer:
n = len(self._reward_buffer)
wandb_metrics["train/mean_reward"] = sum(self._reward_buffer) / n
self._reward_buffer.clear()
await super().wandb_log(wandb_metrics) # MUST call super
```
**Pitfall**: `compute_reward` appends to metric buffers. During eval, this pollutes training metrics. Roll back buffer entries added during eval.
## Config Class
Always create a custom config subclass with Pydantic Field descriptors. Key inherited fields you can tune: `enabled_toolsets`, `max_agent_turns`, `agent_temperature`, `system_prompt`, `terminal_backend`, `group_size`, `steps_per_eval`, `total_steps`.
## config_init() — Default Configuration
Classmethod returning `(YourEnvConfig, [APIServerConfig(...)])`. Set server_type to "openai" for OpenRouter/external APIs. Load API key from environment variable.
## Three CLI Modes
```bash
# SERVE — Full training loop (connects to Atropos API server)
python environments/my_env.py serve --openai.base_url http://localhost:8000/v1
# PROCESS — Offline data generation (saves JSONL)
python environments/my_env.py process --env.total_steps 10 --env.group_size 1 \
--env.use_wandb false --env.data_path_to_save_groups output.jsonl \
--openai.base_url "<USER_BASE_URL>" \
--openai.model_name "<USER_MODEL>" \
--openai.server_type <USER_SERVER_TYPE> --openai.health_check false
# EVALUATE — Standalone eval (runs setup + evaluate only)
python environments/my_env.py evaluate --env.eval_size 20 \
--env.data_dir_to_save_evals /tmp/eval_results \
--openai.base_url "<USER_BASE_URL>" \
--openai.model_name "<USER_MODEL>" \
--openai.server_type <USER_SERVER_TYPE> --openai.health_check false
```
Config priority: CLI args > YAML file > config_init() defaults.
## Common Pitfalls
1. **AgentResult has .messages, not .final_response** — Extract the final response by iterating reversed(result.messages) looking for the last assistant message with content.
2. **evaluate() must use HermesAgentLoop, not chat_completion** — Single-turn chat_completion has no tools. The whole point of hermes-agent benchmarks is agentic evaluation with tool use.
3. **Don't call _llm_judge twice** — If compute_reward already calls it, extract the score from the buffer instead of calling judge separately in evaluate().
4. **Eval pollutes training buffers** — compute_reward appends to metric buffers. During eval, roll back buffer entries to keep training metrics clean.
5. **Always set health_check=false for OpenRouter** — OpenRouter has no /health endpoint.
6. **Set data_dir_to_save_evals in evaluate mode** — Without it, results aren't saved.
7. **default_toolsets class variable vs enabled_toolsets config** — The class variable is a hint; the config field is what actually controls tool resolution.
8. **Tool call parsing in messages** — Tool calls are dicts with `{"function": {"name": ..., "arguments": ...}}`. Always check `isinstance(tc, dict)`.
9. **ToolContext.cleanup()** — Always call in a finally block to release sandbox resources.
10. **server_type must be "openai" for external APIs** — Without it, Atropos assumes a local VLLM server.
11. **Always ask the user for their inference setup** — Never hardcode or assume a specific provider/model. See the "Inference Setup" section above.
## Reward Function Patterns
### LLM Judge (for open-ended tasks)
Use `self.server.chat_completion()` with a scoring prompt. Parse JSON response for score float. Always include a heuristic fallback (keyword overlap) for when the judge call fails.
### Binary Verification (for code/terminal tasks)
Use `ctx.terminal("pytest test.py -q")` to run tests in the agent's sandbox. Return 1.0 for pass, 0.0 for fail.
### Multi-Signal (combine multiple indicators)
Weight correctness (0.6) + tool usage (0.2) + efficiency (0.2) + optional bonuses. Clamp to [0, 1].
## Testing Your Environment
1. **Import test**: `python -c "from environments.my_env import MyEnv; print('OK')"`
2. **Ask the user for inference setup** (see "Inference Setup" section above)
3. **Process mode** (1 item): Verify JSONL output has valid tokens, masks, scores
4. **Evaluate mode**: Verify full agent loop runs with tools, metrics logged correctly
5. **Check reward range**: Scores should be in [0, 1], not all identical
## Minimum Implementation Checklist
```python
class MyEnv(HermesAgentBaseEnv):
name = "my-env"
env_config_cls = MyEnvConfig
@classmethod
def config_init(cls): ... # Default server + env config
async def setup(self): ... # Load dataset + train/eval split
async def get_next_item(self): ... # Cycle through training items
def format_prompt(self, item): ... # Item → user message string
async def compute_reward(self, item, result, ctx): ... # Score rollout
async def evaluate(self, *args, **kwargs): ... # Full agent loop eval
async def wandb_log(self, metrics=None): ... # Custom metrics + super()
if __name__ == "__main__":
MyEnv.cli()
```

534
website/docs/user-guide/skills/optional/mlops/mlops-huggingface-tokenizers.md Normal file
View file

@ -0,0 +1,534 @@
---
title: "Huggingface Tokenizers — Fast tokenizers optimized for research and production"
sidebar_label: "Huggingface Tokenizers"
description: "Fast tokenizers optimized for research and production"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Huggingface Tokenizers
Fast tokenizers optimized for research and production. Rust-based implementation tokenizes 1GB in &lt;20 seconds. Supports BPE, WordPiece, and Unigram algorithms. Train custom vocabularies, track alignments, handle padding/truncation. Integrates seamlessly with transformers. Use when you need high-performance tokenization or custom tokenizer training.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/huggingface-tokenizers` |
| Path | `optional-skills/mlops/huggingface-tokenizers` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `tokenizers`, `transformers`, `datasets` |
| Tags | `Tokenization`, `HuggingFace`, `BPE`, `WordPiece`, `Unigram`, `Fast Tokenization`, `Rust`, `Custom Tokenizer`, `Alignment Tracking`, `Production` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# HuggingFace Tokenizers - Fast Tokenization for NLP
Fast, production-ready tokenizers with Rust performance and Python ease-of-use.
## When to use HuggingFace Tokenizers
**Use HuggingFace Tokenizers when:**
- Need extremely fast tokenization (&lt;20s per GB of text)
- Training custom tokenizers from scratch
- Want alignment tracking (token → original text position)
- Building production NLP pipelines
- Need to tokenize large corpora efficiently
**Performance**:
- **Speed**: &lt;20 seconds to tokenize 1GB on CPU
- **Implementation**: Rust core with Python/Node.js bindings
- **Efficiency**: 10-100× faster than pure Python implementations
**Use alternatives instead**:
- **SentencePiece**: Language-independent, used by T5/ALBERT
- **tiktoken**: OpenAI's BPE tokenizer for GPT models
- **transformers AutoTokenizer**: Loading pretrained only (uses this library internally)
## Quick start
### Installation
```bash
# Install tokenizers
pip install tokenizers
# With transformers integration
pip install tokenizers transformers
```
### Load pretrained tokenizer
```python
from tokenizers import Tokenizer
# Load from HuggingFace Hub
tokenizer = Tokenizer.from_pretrained("bert-base-uncased")
# Encode text
output = tokenizer.encode("Hello, how are you?")
print(output.tokens) # ['hello', ',', 'how', 'are', 'you', '?']
print(output.ids) # [7592, 1010, 2129, 2024, 2017, 1029]
# Decode back
text = tokenizer.decode(output.ids)
print(text) # "hello, how are you?"
```
### Train custom BPE tokenizer
```python
from tokenizers import Tokenizer
from tokenizers.models import BPE
from tokenizers.trainers import BpeTrainer
from tokenizers.pre_tokenizers import Whitespace
# Initialize tokenizer with BPE model
tokenizer = Tokenizer(BPE(unk_token="[UNK]"))
tokenizer.pre_tokenizer = Whitespace()
# Configure trainer
trainer = BpeTrainer(
vocab_size=30000,
special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"],
min_frequency=2
)
# Train on files
files = ["train.txt", "validation.txt"]
tokenizer.train(files, trainer)
# Save
tokenizer.save("my-tokenizer.json")
```
**Training time**: ~1-2 minutes for 100MB corpus, ~10-20 minutes for 1GB
### Batch encoding with padding
```python
# Enable padding
tokenizer.enable_padding(pad_id=3, pad_token="[PAD]")
# Encode batch
texts = ["Hello world", "This is a longer sentence"]
encodings = tokenizer.encode_batch(texts)
for encoding in encodings:
print(encoding.ids)
# [101, 7592, 2088, 102, 3, 3, 3]
# [101, 2023, 2003, 1037, 2936, 6251, 102]
```
## Tokenization algorithms
### BPE (Byte-Pair Encoding)
**How it works**:
1. Start with character-level vocabulary
2. Find most frequent character pair
3. Merge into new token, add to vocabulary
4. Repeat until vocabulary size reached
**Used by**: GPT-2, GPT-3, RoBERTa, BART, DeBERTa
```python
from tokenizers import Tokenizer
from tokenizers.models import BPE
from tokenizers.trainers import BpeTrainer
from tokenizers.pre_tokenizers import ByteLevel
tokenizer = Tokenizer(BPE(unk_token="<|endoftext|>"))
tokenizer.pre_tokenizer = ByteLevel()
trainer = BpeTrainer(
vocab_size=50257,
special_tokens=["<|endoftext|>"],
min_frequency=2
)
tokenizer.train(files=["data.txt"], trainer=trainer)
```
**Advantages**:
- Handles OOV words well (breaks into subwords)
- Flexible vocabulary size
- Good for morphologically rich languages
**Trade-offs**:
- Tokenization depends on merge order
- May split common words unexpectedly
### WordPiece
**How it works**:
1. Start with character vocabulary
2. Score merge pairs: `frequency(pair) / (frequency(first) × frequency(second))`
3. Merge highest scoring pair
4. Repeat until vocabulary size reached
**Used by**: BERT, DistilBERT, MobileBERT
```python
from tokenizers import Tokenizer
from tokenizers.models import WordPiece
from tokenizers.trainers import WordPieceTrainer
from tokenizers.pre_tokenizers import Whitespace
from tokenizers.normalizers import BertNormalizer
tokenizer = Tokenizer(WordPiece(unk_token="[UNK]"))
tokenizer.normalizer = BertNormalizer(lowercase=True)
tokenizer.pre_tokenizer = Whitespace()
trainer = WordPieceTrainer(
vocab_size=30522,
special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"],
continuing_subword_prefix="##"
)
tokenizer.train(files=["corpus.txt"], trainer=trainer)
```
**Advantages**:
- Prioritizes meaningful merges (high score = semantically related)
- Used successfully in BERT (state-of-the-art results)
**Trade-offs**:
- Unknown words become `[UNK]` if no subword match
- Saves vocabulary, not merge rules (larger files)
### Unigram
**How it works**:
1. Start with large vocabulary (all substrings)
2. Compute loss for corpus with current vocabulary
3. Remove tokens with minimal impact on loss
4. Repeat until vocabulary size reached
**Used by**: ALBERT, T5, mBART, XLNet (via SentencePiece)
```python
from tokenizers import Tokenizer
from tokenizers.models import Unigram
from tokenizers.trainers import UnigramTrainer
tokenizer = Tokenizer(Unigram())
trainer = UnigramTrainer(
vocab_size=8000,
special_tokens=["<unk>", "<s>", "</s>"],
unk_token="<unk>"
)
tokenizer.train(files=["data.txt"], trainer=trainer)
```
**Advantages**:
- Probabilistic (finds most likely tokenization)
- Works well for languages without word boundaries
- Handles diverse linguistic contexts
**Trade-offs**:
- Computationally expensive to train
- More hyperparameters to tune
## Tokenization pipeline
Complete pipeline: **Normalization → Pre-tokenization → Model → Post-processing**
### Normalization
Clean and standardize text:
```python
from tokenizers.normalizers import NFD, StripAccents, Lowercase, Sequence
tokenizer.normalizer = Sequence([
NFD(), # Unicode normalization (decompose)
Lowercase(), # Convert to lowercase
StripAccents() # Remove accents
])
# Input: "Héllo WORLD"
# After normalization: "hello world"
```
**Common normalizers**:
- `NFD`, `NFC`, `NFKD`, `NFKC` - Unicode normalization forms
- `Lowercase()` - Convert to lowercase
- `StripAccents()` - Remove accents (é → e)
- `Strip()` - Remove whitespace
- `Replace(pattern, content)` - Regex replacement
### Pre-tokenization
Split text into word-like units:
```python
from tokenizers.pre_tokenizers import Whitespace, Punctuation, Sequence, ByteLevel
# Split on whitespace and punctuation
tokenizer.pre_tokenizer = Sequence([
Whitespace(),
Punctuation()
])
# Input: "Hello, world!"
# After pre-tokenization: ["Hello", ",", "world", "!"]
```
**Common pre-tokenizers**:
- `Whitespace()` - Split on spaces, tabs, newlines
- `ByteLevel()` - GPT-2 style byte-level splitting
- `Punctuation()` - Isolate punctuation
- `Digits(individual_digits=True)` - Split digits individually
- `Metaspace()` - Replace spaces with ▁ (SentencePiece style)
### Post-processing
Add special tokens for model input:
```python
from tokenizers.processors import TemplateProcessing
# BERT-style: [CLS] sentence [SEP]
tokenizer.post_processor = TemplateProcessing(
single="[CLS] $A [SEP]",
pair="[CLS] $A [SEP] $B [SEP]",
special_tokens=[
("[CLS]", 1),
("[SEP]", 2),
],
)
```
**Common patterns**:
```python
# GPT-2: sentence <|endoftext|>
TemplateProcessing(
single="$A <|endoftext|>",
special_tokens=[("<|endoftext|>", 50256)]
)
# RoBERTa: <s> sentence </s>
TemplateProcessing(
single="<s> $A </s>",
pair="<s> $A </s> </s> $B </s>",
special_tokens=[("<s>", 0), ("</s>", 2)]
)
```
## Alignment tracking
Track token positions in original text:
```python
output = tokenizer.encode("Hello, world!")
# Get token offsets
for token, offset in zip(output.tokens, output.offsets):
start, end = offset
print(f"{token:10} → [{start:2}, {end:2}): {text[start:end]!r}")
# Output:
# hello → [ 0, 5): 'Hello'
# , → [ 5, 6): ','
# world → [ 7, 12): 'world'
# ! → [12, 13): '!'
```
**Use cases**:
- Named entity recognition (map predictions back to text)
- Question answering (extract answer spans)
- Token classification (align labels to original positions)
## Integration with transformers
### Load with AutoTokenizer
```python
from transformers import AutoTokenizer
# AutoTokenizer automatically uses fast tokenizers
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
# Check if using fast tokenizer
print(tokenizer.is_fast) # True
# Access underlying tokenizers.Tokenizer
fast_tokenizer = tokenizer.backend_tokenizer
print(type(fast_tokenizer)) # <class 'tokenizers.Tokenizer'>
```
### Convert custom tokenizer to transformers
```python
from tokenizers import Tokenizer
from transformers import PreTrainedTokenizerFast
# Train custom tokenizer
tokenizer = Tokenizer(BPE())
# ... train tokenizer ...
tokenizer.save("my-tokenizer.json")
# Wrap for transformers
transformers_tokenizer = PreTrainedTokenizerFast(
tokenizer_file="my-tokenizer.json",
unk_token="[UNK]",
pad_token="[PAD]",
cls_token="[CLS]",
sep_token="[SEP]",
mask_token="[MASK]"
)
# Use like any transformers tokenizer
outputs = transformers_tokenizer(
"Hello world",
padding=True,
truncation=True,
max_length=512,
return_tensors="pt"
)
```
## Common patterns
### Train from iterator (large datasets)
```python
from datasets import load_dataset
# Load dataset
dataset = load_dataset("wikitext", "wikitext-103-raw-v1", split="train")
# Create batch iterator
def batch_iterator(batch_size=1000):
for i in range(0, len(dataset), batch_size):
yield dataset[i:i + batch_size]["text"]
# Train tokenizer
tokenizer.train_from_iterator(
batch_iterator(),
trainer=trainer,
length=len(dataset) # For progress bar
)
```
**Performance**: Processes 1GB in ~10-20 minutes
### Enable truncation and padding
```python
# Enable truncation
tokenizer.enable_truncation(max_length=512)
# Enable padding
tokenizer.enable_padding(
pad_id=tokenizer.token_to_id("[PAD]"),
pad_token="[PAD]",
length=512 # Fixed length, or None for batch max
)
# Encode with both
output = tokenizer.encode("This is a long sentence that will be truncated...")
print(len(output.ids)) # 512
```
### Multi-processing
```python
from tokenizers import Tokenizer
from multiprocessing import Pool
# Load tokenizer
tokenizer = Tokenizer.from_file("tokenizer.json")
def encode_batch(texts):
return tokenizer.encode_batch(texts)
# Process large corpus in parallel
with Pool(8) as pool:
# Split corpus into chunks
chunk_size = 1000
chunks = [corpus[i:i+chunk_size] for i in range(0, len(corpus), chunk_size)]
# Encode in parallel
results = pool.map(encode_batch, chunks)
```
**Speedup**: 5-8× with 8 cores
## Performance benchmarks
### Training speed
| Corpus Size | BPE (30k vocab) | WordPiece (30k) | Unigram (8k) |
|-------------|-----------------|-----------------|--------------|
| 10 MB | 15 sec | 18 sec | 25 sec |
| 100 MB | 1.5 min | 2 min | 4 min |
| 1 GB | 15 min | 20 min | 40 min |
**Hardware**: 16-core CPU, tested on English Wikipedia
### Tokenization speed
| Implementation | 1 GB corpus | Throughput |
|----------------|-------------|---------------|
| Pure Python | ~20 minutes | ~50 MB/min |
| HF Tokenizers | ~15 seconds | ~4 GB/min |
| **Speedup** | **80×** | **80×** |
**Test**: English text, average sentence length 20 words
### Memory usage
| Task | Memory |
|-------------------------|---------|
| Load tokenizer | ~10 MB |
| Train BPE (30k vocab) | ~200 MB |
| Encode 1M sentences | ~500 MB |
## Supported models
Pre-trained tokenizers available via `from_pretrained()`:
**BERT family**:
- `bert-base-uncased`, `bert-large-cased`
- `distilbert-base-uncased`
- `roberta-base`, `roberta-large`
**GPT family**:
- `gpt2`, `gpt2-medium`, `gpt2-large`
- `distilgpt2`
**T5 family**:
- `t5-small`, `t5-base`, `t5-large`
- `google/flan-t5-xxl`
**Other**:
- `facebook/bart-base`, `facebook/mbart-large-cc25`
- `albert-base-v2`, `albert-xlarge-v2`
- `xlm-roberta-base`, `xlm-roberta-large`
Browse all: https://huggingface.co/models?library=tokenizers
## References
- **[Training Guide](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/huggingface-tokenizers/references/training.md)** - Train custom tokenizers, configure trainers, handle large datasets
- **[Algorithms Deep Dive](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/huggingface-tokenizers/references/algorithms.md)** - BPE, WordPiece, Unigram explained in detail
- **[Pipeline Components](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/huggingface-tokenizers/references/pipeline.md)** - Normalizers, pre-tokenizers, post-processors, decoders
- **[Transformers Integration](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/huggingface-tokenizers/references/integration.md)** - AutoTokenizer, PreTrainedTokenizerFast, special tokens
## Resources
- **Docs**: https://huggingface.co/docs/tokenizers
- **GitHub**: https://github.com/huggingface/tokenizers ⭐ 9,000+
- **Version**: 0.20.0+
- **Course**: https://huggingface.co/learn/nlp-course/chapter6/1
- **Paper**: BPE (Sennrich et al., 2016), WordPiece (Schuster & Nakajima, 2012)

758
website/docs/user-guide/skills/optional/mlops/mlops-instructor.md Normal file
View file

@ -0,0 +1,758 @@
---
title: "Instructor"
sidebar_label: "Instructor"
description: "Extract structured data from LLM responses with Pydantic validation, retry failed extractions automatically, parse complex JSON with type safety, and stream ..."
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Instructor
Extract structured data from LLM responses with Pydantic validation, retry failed extractions automatically, parse complex JSON with type safety, and stream partial results with Instructor - battle-tested structured output library
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/instructor` |
| Path | `optional-skills/mlops/instructor` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `instructor`, `pydantic`, `openai`, `anthropic` |
| Tags | `Prompt Engineering`, `Instructor`, `Structured Output`, `Pydantic`, `Data Extraction`, `JSON Parsing`, `Type Safety`, `Validation`, `Streaming`, `OpenAI`, `Anthropic` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Instructor: Structured LLM Outputs
## When to Use This Skill
Use Instructor when you need to:
- **Extract structured data** from LLM responses reliably
- **Validate outputs** against Pydantic schemas automatically
- **Retry failed extractions** with automatic error handling
- **Parse complex JSON** with type safety and validation
- **Stream partial results** for real-time processing
- **Support multiple LLM providers** with consistent API
**GitHub Stars**: 15,000+ | **Battle-tested**: 100,000+ developers
## Installation
```bash
# Base installation
pip install instructor
# With specific providers
pip install "instructor[anthropic]" # Anthropic Claude
pip install "instructor[openai]" # OpenAI
pip install "instructor[all]" # All providers
```
## Quick Start
### Basic Example: Extract User Data
```python
import instructor
from pydantic import BaseModel
from anthropic import Anthropic
# Define output structure
class User(BaseModel):
name: str
age: int
email: str
# Create instructor client
client = instructor.from_anthropic(Anthropic())
# Extract structured data
user = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": "John Doe is 30 years old. His email is john@example.com"
}],
response_model=User
)
print(user.name) # "John Doe"
print(user.age) # 30
print(user.email) # "john@example.com"
```
### With OpenAI
```python
from openai import OpenAI
client = instructor.from_openai(OpenAI())
user = client.chat.completions.create(
model="gpt-4o-mini",
response_model=User,
messages=[{"role": "user", "content": "Extract: Alice, 25, alice@email.com"}]
)
```
## Core Concepts
### 1. Response Models (Pydantic)
Response models define the structure and validation rules for LLM outputs.
#### Basic Model
```python
from pydantic import BaseModel, Field
class Article(BaseModel):
title: str = Field(description="Article title")
author: str = Field(description="Author name")
word_count: int = Field(description="Number of words", gt=0)
tags: list[str] = Field(description="List of relevant tags")
article = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": "Analyze this article: [article text]"
}],
response_model=Article
)
```
**Benefits:**
- Type safety with Python type hints
- Automatic validation (word_count > 0)
- Self-documenting with Field descriptions
- IDE autocomplete support
#### Nested Models
```python
class Address(BaseModel):
street: str
city: str
country: str
class Person(BaseModel):
name: str
age: int
address: Address # Nested model
person = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": "John lives at 123 Main St, Boston, USA"
}],
response_model=Person
)
print(person.address.city) # "Boston"
```
#### Optional Fields
```python
from typing import Optional
class Product(BaseModel):
name: str
price: float
discount: Optional[float] = None # Optional
description: str = Field(default="No description") # Default value
# LLM doesn't need to provide discount or description
```
#### Enums for Constraints
```python
from enum import Enum
class Sentiment(str, Enum):
POSITIVE = "positive"
NEGATIVE = "negative"
NEUTRAL = "neutral"
class Review(BaseModel):
text: str
sentiment: Sentiment # Only these 3 values allowed
review = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": "This product is amazing!"
}],
response_model=Review
)
print(review.sentiment) # Sentiment.POSITIVE
```
### 2. Validation
Pydantic validates LLM outputs automatically. If validation fails, Instructor retries.
#### Built-in Validators
```python
from pydantic import Field, EmailStr, HttpUrl
class Contact(BaseModel):
name: str = Field(min_length=2, max_length=100)
age: int = Field(ge=0, le=120) # 0 <= age <= 120
email: EmailStr # Validates email format
website: HttpUrl # Validates URL format
# If LLM provides invalid data, Instructor retries automatically
```
#### Custom Validators
```python
from pydantic import field_validator
class Event(BaseModel):
name: str
date: str
attendees: int
@field_validator('date')
def validate_date(cls, v):
"""Ensure date is in YYYY-MM-DD format."""
import re
if not re.match(r'\d{4}-\d{2}-\d{2}', v):
raise ValueError('Date must be YYYY-MM-DD format')
return v
@field_validator('attendees')
def validate_attendees(cls, v):
"""Ensure positive attendees."""
if v < 1:
raise ValueError('Must have at least 1 attendee')
return v
```
#### Model-Level Validation
```python
from pydantic import model_validator
class DateRange(BaseModel):
start_date: str
end_date: str
@model_validator(mode='after')
def check_dates(self):
"""Ensure end_date is after start_date."""
from datetime import datetime
start = datetime.strptime(self.start_date, '%Y-%m-%d')
end = datetime.strptime(self.end_date, '%Y-%m-%d')
if end < start:
raise ValueError('end_date must be after start_date')
return self
```
### 3. Automatic Retrying
Instructor retries automatically when validation fails, providing error feedback to the LLM.
```python
# Retries up to 3 times if validation fails
user = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": "Extract user from: John, age unknown"
}],
response_model=User,
max_retries=3 # Default is 3
)
# If age can't be extracted, Instructor tells the LLM:
# "Validation error: age - field required"
# LLM tries again with better extraction
```
**How it works:**
1. LLM generates output
2. Pydantic validates
3. If invalid: Error message sent back to LLM
4. LLM tries again with error feedback
5. Repeats up to max_retries
### 4. Streaming
Stream partial results for real-time processing.
#### Streaming Partial Objects
```python
from instructor import Partial
class Story(BaseModel):
title: str
content: str
tags: list[str]
# Stream partial updates as LLM generates
for partial_story in client.messages.create_partial(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": "Write a short sci-fi story"
}],
response_model=Story
):
print(f"Title: {partial_story.title}")
print(f"Content so far: {partial_story.content[:100]}...")
# Update UI in real-time
```
#### Streaming Iterables
```python
class Task(BaseModel):
title: str
priority: str
# Stream list items as they're generated
tasks = client.messages.create_iterable(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": "Generate 10 project tasks"
}],
response_model=Task
)
for task in tasks:
print(f"- {task.title} ({task.priority})")
# Process each task as it arrives
```
## Provider Configuration
### Anthropic Claude
```python
import instructor
from anthropic import Anthropic
client = instructor.from_anthropic(
Anthropic(api_key="your-api-key")
)
# Use with Claude models
response = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[...],
response_model=YourModel
)
```
### OpenAI
```python
from openai import OpenAI
client = instructor.from_openai(
OpenAI(api_key="your-api-key")
)
response = client.chat.completions.create(
model="gpt-4o-mini",
response_model=YourModel,
messages=[...]
)
```
### Local Models (Ollama)
```python
from openai import OpenAI
# Point to local Ollama server
client = instructor.from_openai(
OpenAI(
base_url="http://localhost:11434/v1",
api_key="ollama" # Required but ignored
),
mode=instructor.Mode.JSON
)
response = client.chat.completions.create(
model="llama3.1",
response_model=YourModel,
messages=[...]
)
```
## Common Patterns
### Pattern 1: Data Extraction from Text
```python
class CompanyInfo(BaseModel):
name: str
founded_year: int
industry: str
employees: int
headquarters: str
text = """
Tesla, Inc. was founded in 2003. It operates in the automotive and energy
industry with approximately 140,000 employees. The company is headquartered
in Austin, Texas.
"""
company = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": f"Extract company information from: {text}"
}],
response_model=CompanyInfo
)
```
### Pattern 2: Classification
```python
class Category(str, Enum):
TECHNOLOGY = "technology"
FINANCE = "finance"
HEALTHCARE = "healthcare"
EDUCATION = "education"
OTHER = "other"
class ArticleClassification(BaseModel):
category: Category
confidence: float = Field(ge=0.0, le=1.0)
keywords: list[str]
classification = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": "Classify this article: [article text]"
}],
response_model=ArticleClassification
)
```
### Pattern 3: Multi-Entity Extraction
```python
class Person(BaseModel):
name: str
role: str
class Organization(BaseModel):
name: str
industry: str
class Entities(BaseModel):
people: list[Person]
organizations: list[Organization]
locations: list[str]
text = "Tim Cook, CEO of Apple, announced at the event in Cupertino..."
entities = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": f"Extract all entities from: {text}"
}],
response_model=Entities
)
for person in entities.people:
print(f"{person.name} - {person.role}")
```
### Pattern 4: Structured Analysis
```python
class SentimentAnalysis(BaseModel):
overall_sentiment: Sentiment
positive_aspects: list[str]
negative_aspects: list[str]
suggestions: list[str]
score: float = Field(ge=-1.0, le=1.0)
review = "The product works well but setup was confusing..."
analysis = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": f"Analyze this review: {review}"
}],
response_model=SentimentAnalysis
)
```
### Pattern 5: Batch Processing
```python
def extract_person(text: str) -> Person:
return client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": f"Extract person from: {text}"
}],
response_model=Person
)
texts = [
"John Doe is a 30-year-old engineer",
"Jane Smith, 25, works in marketing",
"Bob Johnson, age 40, software developer"
]
people = [extract_person(text) for text in texts]
```
## Advanced Features
### Union Types
```python
from typing import Union
class TextContent(BaseModel):
type: str = "text"
content: str
class ImageContent(BaseModel):
type: str = "image"
url: HttpUrl
caption: str
class Post(BaseModel):
title: str
content: Union[TextContent, ImageContent] # Either type
# LLM chooses appropriate type based on content
```
### Dynamic Models
```python
from pydantic import create_model
# Create model at runtime
DynamicUser = create_model(
'User',
name=(str, ...),
age=(int, Field(ge=0)),
email=(EmailStr, ...)
)
user = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[...],
response_model=DynamicUser
)
```
### Custom Modes
```python
# For providers without native structured outputs
client = instructor.from_anthropic(
Anthropic(),
mode=instructor.Mode.JSON # JSON mode
)
# Available modes:
# - Mode.ANTHROPIC_TOOLS (recommended for Claude)
# - Mode.JSON (fallback)
# - Mode.TOOLS (OpenAI tools)
```
### Context Management
```python
# Single-use client
with instructor.from_anthropic(Anthropic()) as client:
result = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[...],
response_model=YourModel
)
# Client closed automatically
```
## Error Handling
### Handling Validation Errors
```python
from pydantic import ValidationError
try:
user = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[...],
response_model=User,
max_retries=3
)
except ValidationError as e:
print(f"Failed after retries: {e}")
# Handle gracefully
except Exception as e:
print(f"API error: {e}")
```
### Custom Error Messages
```python
class ValidatedUser(BaseModel):
name: str = Field(description="Full name, 2-100 characters")
age: int = Field(description="Age between 0 and 120", ge=0, le=120)
email: EmailStr = Field(description="Valid email address")
class Config:
# Custom error messages
json_schema_extra = {
"examples": [
{
"name": "John Doe",
"age": 30,
"email": "john@example.com"
}
]
}
```
## Best Practices
### 1. Clear Field Descriptions
```python
# ❌ Bad: Vague
class Product(BaseModel):
name: str
price: float
# ✅ Good: Descriptive
class Product(BaseModel):
name: str = Field(description="Product name from the text")
price: float = Field(description="Price in USD, without currency symbol")
```
### 2. Use Appropriate Validation
```python
# ✅ Good: Constrain values
class Rating(BaseModel):
score: int = Field(ge=1, le=5, description="Rating from 1 to 5 stars")
review: str = Field(min_length=10, description="Review text, at least 10 chars")
```
### 3. Provide Examples in Prompts
```python
messages = [{
"role": "user",
"content": """Extract person info from: "John, 30, engineer"
Example format:
{
"name": "John Doe",
"age": 30,
"occupation": "engineer"
}"""
}]
```
### 4. Use Enums for Fixed Categories
```python
# ✅ Good: Enum ensures valid values
class Status(str, Enum):
PENDING = "pending"
APPROVED = "approved"
REJECTED = "rejected"
class Application(BaseModel):
status: Status # LLM must choose from enum
```
### 5. Handle Missing Data Gracefully
```python
class PartialData(BaseModel):
required_field: str
optional_field: Optional[str] = None
default_field: str = "default_value"
# LLM only needs to provide required_field
```
## Comparison to Alternatives
| Feature | Instructor | Manual JSON | LangChain | DSPy |
|---------|------------|-------------|-----------|------|
| Type Safety | ✅ Yes | ❌ No | ⚠️ Partial | ✅ Yes |
| Auto Validation | ✅ Yes | ❌ No | ❌ No | ⚠️ Limited |
| Auto Retry | ✅ Yes | ❌ No | ❌ No | ✅ Yes |
| Streaming | ✅ Yes | ❌ No | ✅ Yes | ❌ No |
| Multi-Provider | ✅ Yes | ⚠️ Manual | ✅ Yes | ✅ Yes |
| Learning Curve | Low | Low | Medium | High |
**When to choose Instructor:**
- Need structured, validated outputs
- Want type safety and IDE support
- Require automatic retries
- Building data extraction systems
**When to choose alternatives:**
- DSPy: Need prompt optimization
- LangChain: Building complex chains
- Manual: Simple, one-off extractions
## Resources
- **Documentation**: https://python.useinstructor.com
- **GitHub**: https://github.com/jxnl/instructor (15k+ stars)
- **Cookbook**: https://python.useinstructor.com/examples
- **Discord**: Community support available
## See Also
- `references/validation.md` - Advanced validation patterns
- `references/providers.md` - Provider-specific configuration
- `references/examples.md` - Real-world use cases

565
website/docs/user-guide/skills/optional/mlops/mlops-lambda-labs.md Normal file
View file

@ -0,0 +1,565 @@
---
title: "Lambda Labs Gpu Cloud — Reserved and on-demand GPU cloud instances for ML training and inference"
sidebar_label: "Lambda Labs Gpu Cloud"
description: "Reserved and on-demand GPU cloud instances for ML training and inference"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Lambda Labs Gpu Cloud
Reserved and on-demand GPU cloud instances for ML training and inference. Use when you need dedicated GPU instances with simple SSH access, persistent filesystems, or high-performance multi-node clusters for large-scale training.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/lambda-labs` |
| Path | `optional-skills/mlops/lambda-labs` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `lambda-cloud-client>=1.0.0` |
| Tags | `Infrastructure`, `GPU Cloud`, `Training`, `Inference`, `Lambda Labs` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Lambda Labs GPU Cloud
Comprehensive guide to running ML workloads on Lambda Labs GPU cloud with on-demand instances and 1-Click Clusters.
## When to use Lambda Labs
**Use Lambda Labs when:**
- Need dedicated GPU instances with full SSH access
- Running long training jobs (hours to days)
- Want simple pricing with no egress fees
- Need persistent storage across sessions
- Require high-performance multi-node clusters (16-512 GPUs)
- Want pre-installed ML stack (Lambda Stack with PyTorch, CUDA, NCCL)
**Key features:**
- **GPU variety**: B200, H100, GH200, A100, A10, A6000, V100
- **Lambda Stack**: Pre-installed PyTorch, TensorFlow, CUDA, cuDNN, NCCL
- **Persistent filesystems**: Keep data across instance restarts
- **1-Click Clusters**: 16-512 GPU Slurm clusters with InfiniBand
- **Simple pricing**: Pay-per-minute, no egress fees
- **Global regions**: 12+ regions worldwide
**Use alternatives instead:**
- **Modal**: For serverless, auto-scaling workloads
- **SkyPilot**: For multi-cloud orchestration and cost optimization
- **RunPod**: For cheaper spot instances and serverless endpoints
- **Vast.ai**: For GPU marketplace with lowest prices
## Quick start
### Account setup
1. Create account at https://lambda.ai
2. Add payment method
3. Generate API key from dashboard
4. Add SSH key (required before launching instances)
### Launch via console
1. Go to https://cloud.lambda.ai/instances
2. Click "Launch instance"
3. Select GPU type and region
4. Choose SSH key
5. Optionally attach filesystem
6. Launch and wait 3-15 minutes
### Connect via SSH
```bash
# Get instance IP from console
ssh ubuntu@<INSTANCE-IP>
# Or with specific key
ssh -i ~/.ssh/lambda_key ubuntu@<INSTANCE-IP>
```
## GPU instances
### Available GPUs
| GPU | VRAM | Price/GPU/hr | Best For |
|-----|------|--------------|----------|
| B200 SXM6 | 180 GB | $4.99 | Largest models, fastest training |
| H100 SXM | 80 GB | $2.99-3.29 | Large model training |
| H100 PCIe | 80 GB | $2.49 | Cost-effective H100 |
| GH200 | 96 GB | $1.49 | Single-GPU large models |
| A100 80GB | 80 GB | $1.79 | Production training |
| A100 40GB | 40 GB | $1.29 | Standard training |
| A10 | 24 GB | $0.75 | Inference, fine-tuning |
| A6000 | 48 GB | $0.80 | Good VRAM/price ratio |
| V100 | 16 GB | $0.55 | Budget training |
### Instance configurations
```
8x GPU: Best for distributed training (DDP, FSDP)
4x GPU: Large models, multi-GPU training
2x GPU: Medium workloads
1x GPU: Fine-tuning, inference, development
```
### Launch times
- Single-GPU: 3-5 minutes
- Multi-GPU: 10-15 minutes
## Lambda Stack
All instances come with Lambda Stack pre-installed:
```bash
# Included software
- Ubuntu 22.04 LTS
- NVIDIA drivers (latest)
- CUDA 12.x
- cuDNN 8.x
- NCCL (for multi-GPU)
- PyTorch (latest)
- TensorFlow (latest)
- JAX
- JupyterLab
```
### Verify installation
```bash
# Check GPU
nvidia-smi
# Check PyTorch
python -c "import torch; print(torch.cuda.is_available())"
# Check CUDA version
nvcc --version
```
## Python API
### Installation
```bash
pip install lambda-cloud-client
```
### Authentication
```python
import os
import lambda_cloud_client
# Configure with API key
configuration = lambda_cloud_client.Configuration(
host="https://cloud.lambdalabs.com/api/v1",
access_token=os.environ["LAMBDA_API_KEY"]
)
```
### List available instances
```python
with lambda_cloud_client.ApiClient(configuration) as api_client:
api = lambda_cloud_client.DefaultApi(api_client)
# Get available instance types
types = api.instance_types()
for name, info in types.data.items():
print(f"{name}: {info.instance_type.description}")
```
### Launch instance
```python
from lambda_cloud_client.models import LaunchInstanceRequest
request = LaunchInstanceRequest(
region_name="us-west-1",
instance_type_name="gpu_1x_h100_sxm5",
ssh_key_names=["my-ssh-key"],
file_system_names=["my-filesystem"], # Optional
name="training-job"
)
response = api.launch_instance(request)
instance_id = response.data.instance_ids[0]
print(f"Launched: {instance_id}")
```
### List running instances
```python
instances = api.list_instances()
for instance in instances.data:
print(f"{instance.name}: {instance.ip} ({instance.status})")
```
### Terminate instance
```python
from lambda_cloud_client.models import TerminateInstanceRequest
request = TerminateInstanceRequest(
instance_ids=[instance_id]
)
api.terminate_instance(request)
```
### SSH key management
```python
from lambda_cloud_client.models import AddSshKeyRequest
# Add SSH key
request = AddSshKeyRequest(
name="my-key",
public_key="ssh-rsa AAAA..."
)
api.add_ssh_key(request)
# List keys
keys = api.list_ssh_keys()
# Delete key
api.delete_ssh_key(key_id)
```
## CLI with curl
### List instance types
```bash
curl -u $LAMBDA_API_KEY: \
https://cloud.lambdalabs.com/api/v1/instance-types | jq
```
### Launch instance
```bash
curl -u $LAMBDA_API_KEY: \
-X POST https://cloud.lambdalabs.com/api/v1/instance-operations/launch \
-H "Content-Type: application/json" \
-d '{
"region_name": "us-west-1",
"instance_type_name": "gpu_1x_h100_sxm5",
"ssh_key_names": ["my-key"]
}' | jq
```
### Terminate instance
```bash
curl -u $LAMBDA_API_KEY: \
-X POST https://cloud.lambdalabs.com/api/v1/instance-operations/terminate \
-H "Content-Type: application/json" \
-d '{"instance_ids": ["<INSTANCE-ID>"]}' | jq
```
## Persistent storage
### Filesystems
Filesystems persist data across instance restarts:
```bash
# Mount location
/lambda/nfs/<FILESYSTEM_NAME>
# Example: save checkpoints
python train.py --checkpoint-dir /lambda/nfs/my-storage/checkpoints
```
### Create filesystem
1. Go to Storage in Lambda console
2. Click "Create filesystem"
3. Select region (must match instance region)
4. Name and create
### Attach to instance
Filesystems must be attached at instance launch time:
- Via console: Select filesystem when launching
- Via API: Include `file_system_names` in launch request
### Best practices
```bash
# Store on filesystem (persists)
/lambda/nfs/storage/
├── datasets/
├── checkpoints/
├── models/
└── outputs/
# Local SSD (faster, ephemeral)
/home/ubuntu/
└── working/ # Temporary files
```
## SSH configuration
### Add SSH key
```bash
# Generate key locally
ssh-keygen -t ed25519 -f ~/.ssh/lambda_key
# Add public key to Lambda console
# Or via API
```
### Multiple keys
```bash
# On instance, add more keys
echo 'ssh-rsa AAAA...' >> ~/.ssh/authorized_keys
```
### Import from GitHub
```bash
# On instance
ssh-import-id gh:username
```
### SSH tunneling
```bash
# Forward Jupyter
ssh -L 8888:localhost:8888 ubuntu@<IP>
# Forward TensorBoard
ssh -L 6006:localhost:6006 ubuntu@<IP>
# Multiple ports
ssh -L 8888:localhost:8888 -L 6006:localhost:6006 ubuntu@<IP>
```
## JupyterLab
### Launch from console
1. Go to Instances page
2. Click "Launch" in Cloud IDE column
3. JupyterLab opens in browser
### Manual access
```bash
# On instance
jupyter lab --ip=0.0.0.0 --port=8888
# From local machine with tunnel
ssh -L 8888:localhost:8888 ubuntu@<IP>
# Open http://localhost:8888
```
## Training workflows
### Single-GPU training
```bash
# SSH to instance
ssh ubuntu@<IP>
# Clone repo
git clone https://github.com/user/project
cd project
# Install dependencies
pip install -r requirements.txt
# Train
python train.py --epochs 100 --checkpoint-dir /lambda/nfs/storage/checkpoints
```
### Multi-GPU training (single node)
```python
# train_ddp.py
import torch
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
def main():
dist.init_process_group("nccl")
rank = dist.get_rank()
device = rank % torch.cuda.device_count()
model = MyModel().to(device)
model = DDP(model, device_ids=[device])
# Training loop...
if __name__ == "__main__":
main()
```
```bash
# Launch with torchrun (8 GPUs)
torchrun --nproc_per_node=8 train_ddp.py
```
### Checkpoint to filesystem
```python
import os
checkpoint_dir = "/lambda/nfs/my-storage/checkpoints"
os.makedirs(checkpoint_dir, exist_ok=True)
# Save checkpoint
torch.save({
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': loss,
}, f"{checkpoint_dir}/checkpoint_{epoch}.pt")
```
## 1-Click Clusters
### Overview
High-performance Slurm clusters with:
- 16-512 NVIDIA H100 or B200 GPUs
- NVIDIA Quantum-2 400 Gb/s InfiniBand
- GPUDirect RDMA at 3200 Gb/s
- Pre-installed distributed ML stack
### Included software
- Ubuntu 22.04 LTS + Lambda Stack
- NCCL, Open MPI
- PyTorch with DDP and FSDP
- TensorFlow
- OFED drivers
### Storage
- 24 TB NVMe per compute node (ephemeral)
- Lambda filesystems for persistent data
### Multi-node training
```bash
# On Slurm cluster
srun --nodes=4 --ntasks-per-node=8 --gpus-per-node=8 \
torchrun --nnodes=4 --nproc_per_node=8 \
--rdzv_backend=c10d --rdzv_endpoint=$MASTER_ADDR:29500 \
train.py
```
## Networking
### Bandwidth
- Inter-instance (same region): up to 200 Gbps
- Internet outbound: 20 Gbps max
### Firewall
- Default: Only port 22 (SSH) open
- Configure additional ports in Lambda console
- ICMP traffic allowed by default
### Private IPs
```bash
# Find private IP
ip addr show | grep 'inet '
```
## Common workflows
### Workflow 1: Fine-tuning LLM
```bash
# 1. Launch 8x H100 instance with filesystem
# 2. SSH and setup
ssh ubuntu@<IP>
pip install transformers accelerate peft
# 3. Download model to filesystem
python -c "
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained('meta-llama/Llama-2-7b-hf')
model.save_pretrained('/lambda/nfs/storage/models/llama-2-7b')
"
# 4. Fine-tune with checkpoints on filesystem
accelerate launch --num_processes 8 train.py \
--model_path /lambda/nfs/storage/models/llama-2-7b \
--output_dir /lambda/nfs/storage/outputs \
--checkpoint_dir /lambda/nfs/storage/checkpoints
```
### Workflow 2: Batch inference
```bash
# 1. Launch A10 instance (cost-effective for inference)
# 2. Run inference
python inference.py \
--model /lambda/nfs/storage/models/fine-tuned \
--input /lambda/nfs/storage/data/inputs.jsonl \
--output /lambda/nfs/storage/data/outputs.jsonl
```
## Cost optimization
### Choose right GPU
| Task | Recommended GPU |
|------|-----------------|
| LLM fine-tuning (7B) | A100 40GB |
| LLM fine-tuning (70B) | 8x H100 |
| Inference | A10, A6000 |
| Development | V100, A10 |
| Maximum performance | B200 |
### Reduce costs
1. **Use filesystems**: Avoid re-downloading data
2. **Checkpoint frequently**: Resume interrupted training
3. **Right-size**: Don't over-provision GPUs
4. **Terminate idle**: No auto-stop, manually terminate
### Monitor usage
- Dashboard shows real-time GPU utilization
- API for programmatic monitoring
## Common issues
| Issue | Solution |
|-------|----------|
| Instance won't launch | Check region availability, try different GPU |
| SSH connection refused | Wait for instance to initialize (3-15 min) |
| Data lost after terminate | Use persistent filesystems |
| Slow data transfer | Use filesystem in same region |
| GPU not detected | Reboot instance, check drivers |
## References
- **[Advanced Usage](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/lambda-labs/references/advanced-usage.md)** - Multi-node training, API automation
- **[Troubleshooting](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/lambda-labs/references/troubleshooting.md)** - Common issues and solutions
## Resources
- **Documentation**: https://docs.lambda.ai
- **Console**: https://cloud.lambda.ai
- **Pricing**: https://lambda.ai/instances
- **Support**: https://support.lambdalabs.com
- **Blog**: https://lambda.ai/blog

322
website/docs/user-guide/skills/optional/mlops/mlops-llava.md Normal file
View file

@ -0,0 +1,322 @@
---
title: "Llava — Large Language and Vision Assistant"
sidebar_label: "Llava"
description: "Large Language and Vision Assistant"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Llava
Large Language and Vision Assistant. Enables visual instruction tuning and image-based conversations. Combines CLIP vision encoder with Vicuna/LLaMA language models. Supports multi-turn image chat, visual question answering, and instruction following. Use for vision-language chatbots or image understanding tasks. Best for conversational image analysis.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/llava` |
| Path | `optional-skills/mlops/llava` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `transformers`, `torch`, `pillow` |
| Tags | `LLaVA`, `Vision-Language`, `Multimodal`, `Visual Question Answering`, `Image Chat`, `CLIP`, `Vicuna`, `Conversational AI`, `Instruction Tuning`, `VQA` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# LLaVA - Large Language and Vision Assistant
Open-source vision-language model for conversational image understanding.
## When to use LLaVA
**Use when:**
- Building vision-language chatbots
- Visual question answering (VQA)
- Image description and captioning
- Multi-turn image conversations
- Visual instruction following
- Document understanding with images
**Metrics**:
- **23,000+ GitHub stars**
- GPT-4V level capabilities (targeted)
- Apache 2.0 License
- Multiple model sizes (7B-34B params)
**Use alternatives instead**:
- **GPT-4V**: Highest quality, API-based
- **CLIP**: Simple zero-shot classification
- **BLIP-2**: Better for captioning only
- **Flamingo**: Research, not open-source
## Quick start
### Installation
```bash
# Clone repository
git clone https://github.com/haotian-liu/LLaVA
cd LLaVA
# Install
pip install -e .
```
### Basic usage
```python
from llava.model.builder import load_pretrained_model
from llava.mm_utils import get_model_name_from_path, process_images, tokenizer_image_token
from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN
from llava.conversation import conv_templates
from PIL import Image
import torch
# Load model
model_path = "liuhaotian/llava-v1.5-7b"
tokenizer, model, image_processor, context_len = load_pretrained_model(
model_path=model_path,
model_base=None,
model_name=get_model_name_from_path(model_path)
)
# Load image
image = Image.open("image.jpg")
image_tensor = process_images([image], image_processor, model.config)
image_tensor = image_tensor.to(model.device, dtype=torch.float16)
# Create conversation
conv = conv_templates["llava_v1"].copy()
conv.append_message(conv.roles[0], DEFAULT_IMAGE_TOKEN + "\nWhat is in this image?")
conv.append_message(conv.roles[1], None)
prompt = conv.get_prompt()
# Generate response
input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).to(model.device)
with torch.inference_mode():
output_ids = model.generate(
input_ids,
images=image_tensor,
do_sample=True,
temperature=0.2,
max_new_tokens=512
)
response = tokenizer.decode(output_ids[0], skip_special_tokens=True).strip()
print(response)
```
## Available models
| Model | Parameters | VRAM | Quality |
|-------|------------|------|---------|
| LLaVA-v1.5-7B | 7B | ~14 GB | Good |
| LLaVA-v1.5-13B | 13B | ~28 GB | Better |
| LLaVA-v1.6-34B | 34B | ~70 GB | Best |
```python
# Load different models
model_7b = "liuhaotian/llava-v1.5-7b"
model_13b = "liuhaotian/llava-v1.5-13b"
model_34b = "liuhaotian/llava-v1.6-34b"
# 4-bit quantization for lower VRAM
load_4bit = True # Reduces VRAM by ~4×
```
## CLI usage
```bash
# Single image query
python -m llava.serve.cli \
--model-path liuhaotian/llava-v1.5-7b \
--image-file image.jpg \
--query "What is in this image?"
# Multi-turn conversation
python -m llava.serve.cli \
--model-path liuhaotian/llava-v1.5-7b \
--image-file image.jpg
# Then type questions interactively
```
## Web UI (Gradio)
```bash
# Launch Gradio interface
python -m llava.serve.gradio_web_server \
--model-path liuhaotian/llava-v1.5-7b \
--load-4bit # Optional: reduce VRAM
# Access at http://localhost:7860
```
## Multi-turn conversations
```python
# Initialize conversation
conv = conv_templates["llava_v1"].copy()
# Turn 1
conv.append_message(conv.roles[0], DEFAULT_IMAGE_TOKEN + "\nWhat is in this image?")
conv.append_message(conv.roles[1], None)
response1 = generate(conv, model, image) # "A dog playing in a park"
# Turn 2
conv.messages[-1][1] = response1 # Add previous response
conv.append_message(conv.roles[0], "What breed is the dog?")
conv.append_message(conv.roles[1], None)
response2 = generate(conv, model, image) # "Golden Retriever"
# Turn 3
conv.messages[-1][1] = response2
conv.append_message(conv.roles[0], "What time of day is it?")
conv.append_message(conv.roles[1], None)
response3 = generate(conv, model, image)
```
## Common tasks
### Image captioning
```python
question = "Describe this image in detail."
response = ask(model, image, question)
```
### Visual question answering
```python
question = "How many people are in the image?"
response = ask(model, image, question)
```
### Object detection (textual)
```python
question = "List all the objects you can see in this image."
response = ask(model, image, question)
```
### Scene understanding
```python
question = "What is happening in this scene?"
response = ask(model, image, question)
```
### Document understanding
```python
question = "What is the main topic of this document?"
response = ask(model, document_image, question)
```
## Training custom model
```bash
# Stage 1: Feature alignment (558K image-caption pairs)
bash scripts/v1_5/pretrain.sh
# Stage 2: Visual instruction tuning (150K instruction data)
bash scripts/v1_5/finetune.sh
```
## Quantization (reduce VRAM)
```python
# 4-bit quantization
tokenizer, model, image_processor, context_len = load_pretrained_model(
model_path="liuhaotian/llava-v1.5-13b",
model_base=None,
model_name=get_model_name_from_path("liuhaotian/llava-v1.5-13b"),
load_4bit=True # Reduces VRAM ~4×
)
# 8-bit quantization
load_8bit=True # Reduces VRAM ~2×
```
## Best practices
1. **Start with 7B model** - Good quality, manageable VRAM
2. **Use 4-bit quantization** - Reduces VRAM significantly
3. **GPU required** - CPU inference extremely slow
4. **Clear prompts** - Specific questions get better answers
5. **Multi-turn conversations** - Maintain conversation context
6. **Temperature 0.2-0.7** - Balance creativity/consistency
7. **max_new_tokens 512-1024** - For detailed responses
8. **Batch processing** - Process multiple images sequentially
## Performance
| Model | VRAM (FP16) | VRAM (4-bit) | Speed (tokens/s) |
|-------|-------------|--------------|------------------|
| 7B | ~14 GB | ~4 GB | ~20 |
| 13B | ~28 GB | ~8 GB | ~12 |
| 34B | ~70 GB | ~18 GB | ~5 |
*On A100 GPU*
## Benchmarks
LLaVA achieves competitive scores on:
- **VQAv2**: 78.5%
- **GQA**: 62.0%
- **MM-Vet**: 35.4%
- **MMBench**: 64.3%
## Limitations
1. **Hallucinations** - May describe things not in image
2. **Spatial reasoning** - Struggles with precise locations
3. **Small text** - Difficulty reading fine print
4. **Object counting** - Imprecise for many objects
5. **VRAM requirements** - Need powerful GPU
6. **Inference speed** - Slower than CLIP
## Integration with frameworks
### LangChain
```python
from langchain.llms.base import LLM
class LLaVALLM(LLM):
def _call(self, prompt, stop=None):
# Custom LLaVA inference
return response
llm = LLaVALLM()
```
### Gradio App
```python
import gradio as gr
def chat(image, text, history):
response = ask_llava(model, image, text)
return response
demo = gr.ChatInterface(
chat,
additional_inputs=[gr.Image(type="pil")],
title="LLaVA Chat"
)
demo.launch()
```
## Resources
- **GitHub**: https://github.com/haotian-liu/LLaVA ⭐ 23,000+
- **Paper**: https://arxiv.org/abs/2304.08485
- **Demo**: https://llava.hliu.cc
- **Models**: https://huggingface.co/liuhaotian
- **License**: Apache 2.0

361
website/docs/user-guide/skills/optional/mlops/mlops-modal.md Normal file
View file

@ -0,0 +1,361 @@
---
title: "Modal Serverless Gpu — Serverless GPU cloud platform for running ML workloads"
sidebar_label: "Modal Serverless Gpu"
description: "Serverless GPU cloud platform for running ML workloads"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Modal Serverless Gpu
Serverless GPU cloud platform for running ML workloads. Use when you need on-demand GPU access without infrastructure management, deploying ML models as APIs, or running batch jobs with automatic scaling.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/modal` |
| Path | `optional-skills/mlops/modal` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `modal>=0.64.0` |
| Tags | `Infrastructure`, `Serverless`, `GPU`, `Cloud`, `Deployment`, `Modal` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Modal Serverless GPU
Comprehensive guide to running ML workloads on Modal's serverless GPU cloud platform.
## When to use Modal
**Use Modal when:**
- Running GPU-intensive ML workloads without managing infrastructure
- Deploying ML models as auto-scaling APIs
- Running batch processing jobs (training, inference, data processing)
- Need pay-per-second GPU pricing without idle costs
- Prototyping ML applications quickly
- Running scheduled jobs (cron-like workloads)
**Key features:**
- **Serverless GPUs**: T4, L4, A10G, L40S, A100, H100, H200, B200 on-demand
- **Python-native**: Define infrastructure in Python code, no YAML
- **Auto-scaling**: Scale to zero, scale to 100+ GPUs instantly
- **Sub-second cold starts**: Rust-based infrastructure for fast container launches
- **Container caching**: Image layers cached for rapid iteration
- **Web endpoints**: Deploy functions as REST APIs with zero-downtime updates
**Use alternatives instead:**
- **RunPod**: For longer-running pods with persistent state
- **Lambda Labs**: For reserved GPU instances
- **SkyPilot**: For multi-cloud orchestration and cost optimization
- **Kubernetes**: For complex multi-service architectures
## Quick start
### Installation
```bash
pip install modal
modal setup # Opens browser for authentication
```
### Hello World with GPU
```python
import modal
app = modal.App("hello-gpu")
@app.function(gpu="T4")
def gpu_info():
import subprocess
return subprocess.run(["nvidia-smi"], capture_output=True, text=True).stdout
@app.local_entrypoint()
def main():
print(gpu_info.remote())
```
Run: `modal run hello_gpu.py`
### Basic inference endpoint
```python
import modal
app = modal.App("text-generation")
image = modal.Image.debian_slim().pip_install("transformers", "torch", "accelerate")
@app.cls(gpu="A10G", image=image)
class TextGenerator:
@modal.enter()
def load_model(self):
from transformers import pipeline
self.pipe = pipeline("text-generation", model="gpt2", device=0)
@modal.method()
def generate(self, prompt: str) -> str:
return self.pipe(prompt, max_length=100)[0]["generated_text"]
@app.local_entrypoint()
def main():
print(TextGenerator().generate.remote("Hello, world"))
```
## Core concepts
### Key components
| Component | Purpose |
|-----------|---------|
| `App` | Container for functions and resources |
| `Function` | Serverless function with compute specs |
| `Cls` | Class-based functions with lifecycle hooks |
| `Image` | Container image definition |
| `Volume` | Persistent storage for models/data |
| `Secret` | Secure credential storage |
### Execution modes
| Command | Description |
|---------|-------------|
| `modal run script.py` | Execute and exit |
| `modal serve script.py` | Development with live reload |
| `modal deploy script.py` | Persistent cloud deployment |
## GPU configuration
### Available GPUs
| GPU | VRAM | Best For |
|-----|------|----------|
| `T4` | 16GB | Budget inference, small models |
| `L4` | 24GB | Inference, Ada Lovelace arch |
| `A10G` | 24GB | Training/inference, 3.3x faster than T4 |
| `L40S` | 48GB | Recommended for inference (best cost/perf) |
| `A100-40GB` | 40GB | Large model training |
| `A100-80GB` | 80GB | Very large models |
| `H100` | 80GB | Fastest, FP8 + Transformer Engine |
| `H200` | 141GB | Auto-upgrade from H100, 4.8TB/s bandwidth |
| `B200` | Latest | Blackwell architecture |
### GPU specification patterns
```python
# Single GPU
@app.function(gpu="A100")
# Specific memory variant
@app.function(gpu="A100-80GB")
# Multiple GPUs (up to 8)
@app.function(gpu="H100:4")
# GPU with fallbacks
@app.function(gpu=["H100", "A100", "L40S"])
# Any available GPU
@app.function(gpu="any")
```
## Container images
```python
# Basic image with pip
image = modal.Image.debian_slim(python_version="3.11").pip_install(
"torch==2.1.0", "transformers==4.36.0", "accelerate"
)
# From CUDA base
image = modal.Image.from_registry(
"nvidia/cuda:12.1.0-cudnn8-devel-ubuntu22.04",
add_python="3.11"
).pip_install("torch", "transformers")
# With system packages
image = modal.Image.debian_slim().apt_install("git", "ffmpeg").pip_install("whisper")
```
## Persistent storage
```python
volume = modal.Volume.from_name("model-cache", create_if_missing=True)
@app.function(gpu="A10G", volumes={"/models": volume})
def load_model():
import os
model_path = "/models/llama-7b"
if not os.path.exists(model_path):
model = download_model()
model.save_pretrained(model_path)
volume.commit() # Persist changes
return load_from_path(model_path)
```
## Web endpoints
### FastAPI endpoint decorator
```python
@app.function()
@modal.fastapi_endpoint(method="POST")
def predict(text: str) -> dict:
return {"result": model.predict(text)}
```
### Full ASGI app
```python
from fastapi import FastAPI
web_app = FastAPI()
@web_app.post("/predict")
async def predict(text: str):
return {"result": await model.predict.remote.aio(text)}
@app.function()
@modal.asgi_app()
def fastapi_app():
return web_app
```
### Web endpoint types
| Decorator | Use Case |
|-----------|----------|
| `@modal.fastapi_endpoint()` | Simple function → API |
| `@modal.asgi_app()` | Full FastAPI/Starlette apps |
| `@modal.wsgi_app()` | Django/Flask apps |
| `@modal.web_server(port)` | Arbitrary HTTP servers |
## Dynamic batching
```python
@app.function()
@modal.batched(max_batch_size=32, wait_ms=100)
async def batch_predict(inputs: list[str]) -> list[dict]:
# Inputs automatically batched
return model.batch_predict(inputs)
```
## Secrets management
```bash
# Create secret
modal secret create huggingface HF_TOKEN=hf_xxx
```
```python
@app.function(secrets=[modal.Secret.from_name("huggingface")])
def download_model():
import os
token = os.environ["HF_TOKEN"]
```
## Scheduling
```python
@app.function(schedule=modal.Cron("0 0 * * *")) # Daily midnight
def daily_job():
pass
@app.function(schedule=modal.Period(hours=1))
def hourly_job():
pass
```
## Performance optimization
### Cold start mitigation
```python
@app.function(
container_idle_timeout=300, # Keep warm 5 min
allow_concurrent_inputs=10, # Handle concurrent requests
)
def inference():
pass
```
### Model loading best practices
```python
@app.cls(gpu="A100")
class Model:
@modal.enter() # Run once at container start
def load(self):
self.model = load_model() # Load during warm-up
@modal.method()
def predict(self, x):
return self.model(x)
```
## Parallel processing
```python
@app.function()
def process_item(item):
return expensive_computation(item)
@app.function()
def run_parallel():
items = list(range(1000))
# Fan out to parallel containers
results = list(process_item.map(items))
return results
```
## Common configuration
```python
@app.function(
gpu="A100",
memory=32768, # 32GB RAM
cpu=4, # 4 CPU cores
timeout=3600, # 1 hour max
container_idle_timeout=120,# Keep warm 2 min
retries=3, # Retry on failure
concurrency_limit=10, # Max concurrent containers
)
def my_function():
pass
```
## Debugging
```python
# Test locally
if __name__ == "__main__":
result = my_function.local()
# View logs
# modal app logs my-app
```
## Common issues
| Issue | Solution |
|-------|----------|
| Cold start latency | Increase `container_idle_timeout`, use `@modal.enter()` |
| GPU OOM | Use larger GPU (`A100-80GB`), enable gradient checkpointing |
| Image build fails | Pin dependency versions, check CUDA compatibility |
| Timeout errors | Increase `timeout`, add checkpointing |
## References
- **[Advanced Usage](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/modal/references/advanced-usage.md)** - Multi-GPU, distributed training, cost optimization
- **[Troubleshooting](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/modal/references/troubleshooting.md)** - Common issues and solutions
## Resources
- **Documentation**: https://modal.com/docs
- **Examples**: https://github.com/modal-labs/modal-examples
- **Pricing**: https://modal.com/pricing
- **Discord**: https://discord.gg/modal

400
website/docs/user-guide/skills/optional/mlops/mlops-nemo-curator.md Normal file
View file

@ -0,0 +1,400 @@
---
title: "Nemo Curator — GPU-accelerated data curation for LLM training"
sidebar_label: "Nemo Curator"
description: "GPU-accelerated data curation for LLM training"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Nemo Curator
GPU-accelerated data curation for LLM training. Supports text/image/video/audio. Features fuzzy deduplication (16× faster), quality filtering (30+ heuristics), semantic deduplication, PII redaction, NSFW detection. Scales across GPUs with RAPIDS. Use for preparing high-quality training datasets, cleaning web data, or deduplicating large corpora.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/nemo-curator` |
| Path | `optional-skills/mlops/nemo-curator` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `nemo-curator`, `cudf`, `dask`, `rapids` |
| Tags | `Data Processing`, `NeMo Curator`, `Data Curation`, `GPU Acceleration`, `Deduplication`, `Quality Filtering`, `NVIDIA`, `RAPIDS`, `PII Redaction`, `Multimodal`, `LLM Training Data` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# NeMo Curator - GPU-Accelerated Data Curation
NVIDIA's toolkit for preparing high-quality training data for LLMs.
## When to use NeMo Curator
**Use NeMo Curator when:**
- Preparing LLM training data from web scrapes (Common Crawl)
- Need fast deduplication (16× faster than CPU)
- Curating multi-modal datasets (text, images, video, audio)
- Filtering low-quality or toxic content
- Scaling data processing across GPU cluster
**Performance**:
- **16× faster** fuzzy deduplication (8TB RedPajama v2)
- **40% lower TCO** vs CPU alternatives
- **Near-linear scaling** across GPU nodes
**Use alternatives instead**:
- **datatrove**: CPU-based, open-source data processing
- **dolma**: Allen AI's data toolkit
- **Ray Data**: General ML data processing (no curation focus)
## Quick start
### Installation
```bash
# Text curation (CUDA 12)
uv pip install "nemo-curator[text_cuda12]"
# All modalities
uv pip install "nemo-curator[all_cuda12]"
# CPU-only (slower)
uv pip install "nemo-curator[cpu]"
```
### Basic text curation pipeline
```python
from nemo_curator import ScoreFilter, Modify
from nemo_curator.datasets import DocumentDataset
import pandas as pd
# Load data
df = pd.DataFrame({"text": ["Good document", "Bad doc", "Excellent text"]})
dataset = DocumentDataset(df)
# Quality filtering
def quality_score(doc):
return len(doc["text"].split()) > 5 # Filter short docs
filtered = ScoreFilter(quality_score)(dataset)
# Deduplication
from nemo_curator.modules import ExactDuplicates
deduped = ExactDuplicates()(filtered)
# Save
deduped.to_parquet("curated_data/")
```
## Data curation pipeline
### Stage 1: Quality filtering
```python
from nemo_curator.filters import (
WordCountFilter,
RepeatedLinesFilter,
UrlRatioFilter,
NonAlphaNumericFilter
)
# Apply 30+ heuristic filters
from nemo_curator import ScoreFilter
# Word count filter
dataset = dataset.filter(WordCountFilter(min_words=50, max_words=100000))
# Remove repetitive content
dataset = dataset.filter(RepeatedLinesFilter(max_repeated_line_fraction=0.3))
# URL ratio filter
dataset = dataset.filter(UrlRatioFilter(max_url_ratio=0.2))
```
### Stage 2: Deduplication
**Exact deduplication**:
```python
from nemo_curator.modules import ExactDuplicates
# Remove exact duplicates
deduped = ExactDuplicates(id_field="id", text_field="text")(dataset)
```
**Fuzzy deduplication** (16× faster on GPU):
```python
from nemo_curator.modules import FuzzyDuplicates
# MinHash + LSH deduplication
fuzzy_dedup = FuzzyDuplicates(
id_field="id",
text_field="text",
num_hashes=260, # MinHash parameters
num_buckets=20,
hash_method="md5"
)
deduped = fuzzy_dedup(dataset)
```
**Semantic deduplication**:
```python
from nemo_curator.modules import SemanticDuplicates
# Embedding-based deduplication
semantic_dedup = SemanticDuplicates(
id_field="id",
text_field="text",
embedding_model="sentence-transformers/all-MiniLM-L6-v2",
threshold=0.8 # Cosine similarity threshold
)
deduped = semantic_dedup(dataset)
```
### Stage 3: PII redaction
```python
from nemo_curator.modules import Modify
from nemo_curator.modifiers import PIIRedactor
# Redact personally identifiable information
pii_redactor = PIIRedactor(
supported_entities=["EMAIL_ADDRESS", "PHONE_NUMBER", "PERSON", "LOCATION"],
anonymize_action="replace" # or "redact"
)
redacted = Modify(pii_redactor)(dataset)
```
### Stage 4: Classifier filtering
```python
from nemo_curator.classifiers import QualityClassifier
# Quality classification
quality_clf = QualityClassifier(
model_path="nvidia/quality-classifier-deberta",
batch_size=256,
device="cuda"
)
# Filter low-quality documents
high_quality = dataset.filter(lambda doc: quality_clf(doc["text"]) > 0.5)
```
## GPU acceleration
### GPU vs CPU performance
| Operation | CPU (16 cores) | GPU (A100) | Speedup |
|-----------|----------------|------------|---------|
| Fuzzy dedup (8TB) | 120 hours | 7.5 hours | 16× |
| Exact dedup (1TB) | 8 hours | 0.5 hours | 16× |
| Quality filtering | 2 hours | 0.2 hours | 10× |
### Multi-GPU scaling
```python
from nemo_curator import get_client
import dask_cuda
# Initialize GPU cluster
client = get_client(cluster_type="gpu", n_workers=8)
# Process with 8 GPUs
deduped = FuzzyDuplicates(...)(dataset)
```
## Multi-modal curation
### Image curation
```python
from nemo_curator.image import (
AestheticFilter,
NSFWFilter,
CLIPEmbedder
)
# Aesthetic scoring
aesthetic_filter = AestheticFilter(threshold=5.0)
filtered_images = aesthetic_filter(image_dataset)
# NSFW detection
nsfw_filter = NSFWFilter(threshold=0.9)
safe_images = nsfw_filter(filtered_images)
# Generate CLIP embeddings
clip_embedder = CLIPEmbedder(model="openai/clip-vit-base-patch32")
image_embeddings = clip_embedder(safe_images)
```
### Video curation
```python
from nemo_curator.video import (
SceneDetector,
ClipExtractor,
InternVideo2Embedder
)
# Detect scenes
scene_detector = SceneDetector(threshold=27.0)
scenes = scene_detector(video_dataset)
# Extract clips
clip_extractor = ClipExtractor(min_duration=2.0, max_duration=10.0)
clips = clip_extractor(scenes)
# Generate embeddings
video_embedder = InternVideo2Embedder()
video_embeddings = video_embedder(clips)
```
### Audio curation
```python
from nemo_curator.audio import (
ASRInference,
WERFilter,
DurationFilter
)
# ASR transcription
asr = ASRInference(model="nvidia/stt_en_fastconformer_hybrid_large_pc")
transcribed = asr(audio_dataset)
# Filter by WER (word error rate)
wer_filter = WERFilter(max_wer=0.3)
high_quality_audio = wer_filter(transcribed)
# Duration filtering
duration_filter = DurationFilter(min_duration=1.0, max_duration=30.0)
filtered_audio = duration_filter(high_quality_audio)
```
## Common patterns
### Web scrape curation (Common Crawl)
```python
from nemo_curator import ScoreFilter, Modify
from nemo_curator.filters import *
from nemo_curator.modules import *
from nemo_curator.datasets import DocumentDataset
# Load Common Crawl data
dataset = DocumentDataset.read_parquet("common_crawl/*.parquet")
# Pipeline
pipeline = [
# 1. Quality filtering
WordCountFilter(min_words=100, max_words=50000),
RepeatedLinesFilter(max_repeated_line_fraction=0.2),
SymbolToWordRatioFilter(max_symbol_to_word_ratio=0.3),
UrlRatioFilter(max_url_ratio=0.3),
# 2. Language filtering
LanguageIdentificationFilter(target_languages=["en"]),
# 3. Deduplication
ExactDuplicates(id_field="id", text_field="text"),
FuzzyDuplicates(id_field="id", text_field="text", num_hashes=260),
# 4. PII redaction
PIIRedactor(),
# 5. NSFW filtering
NSFWClassifier(threshold=0.8)
]
# Execute
for stage in pipeline:
dataset = stage(dataset)
# Save
dataset.to_parquet("curated_common_crawl/")
```
### Distributed processing
```python
from nemo_curator import get_client
from dask_cuda import LocalCUDACluster
# Multi-GPU cluster
cluster = LocalCUDACluster(n_workers=8)
client = get_client(cluster=cluster)
# Process large dataset
dataset = DocumentDataset.read_parquet("s3://large_dataset/*.parquet")
deduped = FuzzyDuplicates(...)(dataset)
# Cleanup
client.close()
cluster.close()
```
## Performance benchmarks
### Fuzzy deduplication (8TB RedPajama v2)
- **CPU (256 cores)**: 120 hours
- **GPU (8× A100)**: 7.5 hours
- **Speedup**: 16×
### Exact deduplication (1TB)
- **CPU (64 cores)**: 8 hours
- **GPU (4× A100)**: 0.5 hours
- **Speedup**: 16×
### Quality filtering (100GB)
- **CPU (32 cores)**: 2 hours
- **GPU (2× A100)**: 0.2 hours
- **Speedup**: 10×
## Cost comparison
**CPU-based curation** (AWS c5.18xlarge × 10):
- Cost: $3.60/hour × 10 = $36/hour
- Time for 8TB: 120 hours
- **Total**: $4,320
**GPU-based curation** (AWS p4d.24xlarge × 2):
- Cost: $32.77/hour × 2 = $65.54/hour
- Time for 8TB: 7.5 hours
- **Total**: $491.55
**Savings**: 89% reduction ($3,828 saved)
## Supported data formats
- **Input**: Parquet, JSONL, CSV
- **Output**: Parquet (recommended), JSONL
- **WebDataset**: TAR archives for multi-modal
## Use cases
**Production deployments**:
- NVIDIA used NeMo Curator to prepare Nemotron-4 training data
- Open-source datasets curated: RedPajama v2, The Pile
## References
- **[Filtering Guide](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/nemo-curator/references/filtering.md)** - 30+ quality filters, heuristics
- **[Deduplication Guide](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/nemo-curator/references/deduplication.md)** - Exact, fuzzy, semantic methods
## Resources
- **GitHub**: https://github.com/NVIDIA/NeMo-Curator ⭐ 500+
- **Docs**: https://docs.nvidia.com/nemo-framework/user-guide/latest/datacuration/
- **Version**: 0.4.0+
- **License**: Apache 2.0

451
website/docs/user-guide/skills/optional/mlops/mlops-peft.md Normal file
View file

@ -0,0 +1,451 @@
---
title: "Peft Fine Tuning — Parameter-efficient fine-tuning for LLMs using LoRA, QLoRA, and 25+ methods"
sidebar_label: "Peft Fine Tuning"
description: "Parameter-efficient fine-tuning for LLMs using LoRA, QLoRA, and 25+ methods"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Peft Fine Tuning
Parameter-efficient fine-tuning for LLMs using LoRA, QLoRA, and 25+ methods. Use when fine-tuning large models (7B-70B) with limited GPU memory, when you need to train &lt;1% of parameters with minimal accuracy loss, or for multi-adapter serving. HuggingFace's official library integrated with transformers ecosystem.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/peft` |
| Path | `optional-skills/mlops/peft` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `peft>=0.13.0`, `transformers>=4.45.0`, `torch>=2.0.0`, `bitsandbytes>=0.43.0` |
| Tags | `Fine-Tuning`, `PEFT`, `LoRA`, `QLoRA`, `Parameter-Efficient`, `Adapters`, `Low-Rank`, `Memory Optimization`, `Multi-Adapter` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# PEFT (Parameter-Efficient Fine-Tuning)
Fine-tune LLMs by training &lt;1% of parameters using LoRA, QLoRA, and 25+ adapter methods.
## When to use PEFT
**Use PEFT/LoRA when:**
- Fine-tuning 7B-70B models on consumer GPUs (RTX 4090, A100)
- Need to train &lt;1% parameters (6MB adapters vs 14GB full model)
- Want fast iteration with multiple task-specific adapters
- Deploying multiple fine-tuned variants from one base model
**Use QLoRA (PEFT + quantization) when:**
- Fine-tuning 70B models on single 24GB GPU
- Memory is the primary constraint
- Can accept ~5% quality trade-off vs full fine-tuning
**Use full fine-tuning instead when:**
- Training small models (&lt;1B parameters)
- Need maximum quality and have compute budget
- Significant domain shift requires updating all weights
## Quick start
### Installation
```bash
# Basic installation
pip install peft
# With quantization support (recommended)
pip install peft bitsandbytes
# Full stack
pip install peft transformers accelerate bitsandbytes datasets
```
### LoRA fine-tuning (standard)
```python
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer
from peft import get_peft_model, LoraConfig, TaskType
from datasets import load_dataset
# Load base model
model_name = "meta-llama/Llama-3.1-8B"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.pad_token = tokenizer.eos_token
# LoRA configuration
lora_config = LoraConfig(
task_type=TaskType.CAUSAL_LM,
r=16, # Rank (8-64, higher = more capacity)
lora_alpha=32, # Scaling factor (typically 2*r)
lora_dropout=0.05, # Dropout for regularization
target_modules=["q_proj", "v_proj", "k_proj", "o_proj"], # Attention layers
bias="none" # Don't train biases
)
# Apply LoRA
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# Output: trainable params: 13,631,488 || all params: 8,043,307,008 || trainable%: 0.17%
# Prepare dataset
dataset = load_dataset("databricks/databricks-dolly-15k", split="train")
def tokenize(example):
text = f"### Instruction:\n{example['instruction']}\n\n### Response:\n{example['response']}"
return tokenizer(text, truncation=True, max_length=512, padding="max_length")
tokenized = dataset.map(tokenize, remove_columns=dataset.column_names)
# Training
training_args = TrainingArguments(
output_dir="./lora-llama",
num_train_epochs=3,
per_device_train_batch_size=4,
gradient_accumulation_steps=4,
learning_rate=2e-4,
fp16=True,
logging_steps=10,
save_strategy="epoch"
)
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized,
data_collator=lambda data: {"input_ids": torch.stack([f["input_ids"] for f in data]),
"attention_mask": torch.stack([f["attention_mask"] for f in data]),
"labels": torch.stack([f["input_ids"] for f in data])}
)
trainer.train()
# Save adapter only (6MB vs 16GB)
model.save_pretrained("./lora-llama-adapter")
```
### QLoRA fine-tuning (memory-efficient)
```python
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
from peft import get_peft_model, LoraConfig, prepare_model_for_kbit_training
# 4-bit quantization config
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4", # NormalFloat4 (best for LLMs)
bnb_4bit_compute_dtype="bfloat16", # Compute in bf16
bnb_4bit_use_double_quant=True # Nested quantization
)
# Load quantized model
model = AutoModelForCausalLM.from_pretrained(
"meta-llama/Llama-3.1-70B",
quantization_config=bnb_config,
device_map="auto"
)
# Prepare for training (enables gradient checkpointing)
model = prepare_model_for_kbit_training(model)
# LoRA config for QLoRA
lora_config = LoraConfig(
r=64, # Higher rank for 70B
lora_alpha=128,
lora_dropout=0.1,
target_modules=["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
bias="none",
task_type="CAUSAL_LM"
)
model = get_peft_model(model, lora_config)
# 70B model now fits on single 24GB GPU!
```
## LoRA parameter selection
### Rank (r) - capacity vs efficiency
| Rank | Trainable Params | Memory | Quality | Use Case |
|------|-----------------|--------|---------|----------|
| 4 | ~3M | Minimal | Lower | Simple tasks, prototyping |
| **8** | ~7M | Low | Good | **Recommended starting point** |
| **16** | ~14M | Medium | Better | **General fine-tuning** |
| 32 | ~27M | Higher | High | Complex tasks |
| 64 | ~54M | High | Highest | Domain adaptation, 70B models |
### Alpha (lora_alpha) - scaling factor
```python
# Rule of thumb: alpha = 2 * rank
LoraConfig(r=16, lora_alpha=32) # Standard
LoraConfig(r=16, lora_alpha=16) # Conservative (lower learning rate effect)
LoraConfig(r=16, lora_alpha=64) # Aggressive (higher learning rate effect)
```
### Target modules by architecture
```python
# Llama / Mistral / Qwen
target_modules = ["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
# GPT-2 / GPT-Neo
target_modules = ["c_attn", "c_proj", "c_fc"]
# Falcon
target_modules = ["query_key_value", "dense", "dense_h_to_4h", "dense_4h_to_h"]
# BLOOM
target_modules = ["query_key_value", "dense", "dense_h_to_4h", "dense_4h_to_h"]
# Auto-detect all linear layers
target_modules = "all-linear" # PEFT 0.6.0+
```
## Loading and merging adapters
### Load trained adapter
```python
from peft import PeftModel, AutoPeftModelForCausalLM
from transformers import AutoModelForCausalLM
# Option 1: Load with PeftModel
base_model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B")
model = PeftModel.from_pretrained(base_model, "./lora-llama-adapter")
# Option 2: Load directly (recommended)
model = AutoPeftModelForCausalLM.from_pretrained(
"./lora-llama-adapter",
device_map="auto"
)
```
### Merge adapter into base model
```python
# Merge for deployment (no adapter overhead)
merged_model = model.merge_and_unload()
# Save merged model
merged_model.save_pretrained("./llama-merged")
tokenizer.save_pretrained("./llama-merged")
# Push to Hub
merged_model.push_to_hub("username/llama-finetuned")
```
### Multi-adapter serving
```python
from peft import PeftModel
# Load base with first adapter
model = AutoPeftModelForCausalLM.from_pretrained("./adapter-task1")
# Load additional adapters
model.load_adapter("./adapter-task2", adapter_name="task2")
model.load_adapter("./adapter-task3", adapter_name="task3")
# Switch between adapters at runtime
model.set_adapter("task1") # Use task1 adapter
output1 = model.generate(**inputs)
model.set_adapter("task2") # Switch to task2
output2 = model.generate(**inputs)
# Disable adapters (use base model)
with model.disable_adapter():
base_output = model.generate(**inputs)
```
## PEFT methods comparison
| Method | Trainable % | Memory | Speed | Best For |
|--------|------------|--------|-------|----------|
| **LoRA** | 0.1-1% | Low | Fast | General fine-tuning |
| **QLoRA** | 0.1-1% | Very Low | Medium | Memory-constrained |
| AdaLoRA | 0.1-1% | Low | Medium | Automatic rank selection |
| IA3 | 0.01% | Minimal | Fastest | Few-shot adaptation |
| Prefix Tuning | 0.1% | Low | Medium | Generation control |
| Prompt Tuning | 0.001% | Minimal | Fast | Simple task adaptation |
| P-Tuning v2 | 0.1% | Low | Medium | NLU tasks |
### IA3 (minimal parameters)
```python
from peft import IA3Config
ia3_config = IA3Config(
target_modules=["q_proj", "v_proj", "k_proj", "down_proj"],
feedforward_modules=["down_proj"]
)
model = get_peft_model(model, ia3_config)
# Trains only 0.01% of parameters!
```
### Prefix Tuning
```python
from peft import PrefixTuningConfig
prefix_config = PrefixTuningConfig(
task_type="CAUSAL_LM",
num_virtual_tokens=20, # Prepended tokens
prefix_projection=True # Use MLP projection
)
model = get_peft_model(model, prefix_config)
```
## Integration patterns
### With TRL (SFTTrainer)
```python
from trl import SFTTrainer, SFTConfig
from peft import LoraConfig
lora_config = LoraConfig(r=16, lora_alpha=32, target_modules="all-linear")
trainer = SFTTrainer(
model=model,
args=SFTConfig(output_dir="./output", max_seq_length=512),
train_dataset=dataset,
peft_config=lora_config, # Pass LoRA config directly
)
trainer.train()
```
### With Axolotl (YAML config)
```yaml
# axolotl config.yaml
adapter: lora
lora_r: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target_modules:
- q_proj
- v_proj
- k_proj
- o_proj
lora_target_linear: true # Target all linear layers
```
### With vLLM (inference)
```python
from vllm import LLM
from vllm.lora.request import LoRARequest
# Load base model with LoRA support
llm = LLM(model="meta-llama/Llama-3.1-8B", enable_lora=True)
# Serve with adapter
outputs = llm.generate(
prompts,
lora_request=LoRARequest("adapter1", 1, "./lora-adapter")
)
```
## Performance benchmarks
### Memory usage (Llama 3.1 8B)
| Method | GPU Memory | Trainable Params |
|--------|-----------|------------------|
| Full fine-tuning | 60+ GB | 8B (100%) |
| LoRA r=16 | 18 GB | 14M (0.17%) |
| QLoRA r=16 | 6 GB | 14M (0.17%) |
| IA3 | 16 GB | 800K (0.01%) |
### Training speed (A100 80GB)
| Method | Tokens/sec | vs Full FT |
|--------|-----------|------------|
| Full FT | 2,500 | 1x |
| LoRA | 3,200 | 1.3x |
| QLoRA | 2,100 | 0.84x |
### Quality (MMLU benchmark)
| Model | Full FT | LoRA | QLoRA |
|-------|---------|------|-------|
| Llama 2-7B | 45.3 | 44.8 | 44.1 |
| Llama 2-13B | 54.8 | 54.2 | 53.5 |
## Common issues
### CUDA OOM during training
```python
# Solution 1: Enable gradient checkpointing
model.gradient_checkpointing_enable()
# Solution 2: Reduce batch size + increase accumulation
TrainingArguments(
per_device_train_batch_size=1,
gradient_accumulation_steps=16
)
# Solution 3: Use QLoRA
from transformers import BitsAndBytesConfig
bnb_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4")
```
### Adapter not applying
```python
# Verify adapter is active
print(model.active_adapters) # Should show adapter name
# Check trainable parameters
model.print_trainable_parameters()
# Ensure model in training mode
model.train()
```
### Quality degradation
```python
# Increase rank
LoraConfig(r=32, lora_alpha=64)
# Target more modules
target_modules = "all-linear"
# Use more training data and epochs
TrainingArguments(num_train_epochs=5)
# Lower learning rate
TrainingArguments(learning_rate=1e-4)
```
## Best practices
1. **Start with r=8-16**, increase if quality insufficient
2. **Use alpha = 2 * rank** as starting point
3. **Target attention + MLP layers** for best quality/efficiency
4. **Enable gradient checkpointing** for memory savings
5. **Save adapters frequently** (small files, easy rollback)
6. **Evaluate on held-out data** before merging
7. **Use QLoRA for 70B+ models** on consumer hardware
## References
- **[Advanced Usage](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/peft/references/advanced-usage.md)** - DoRA, LoftQ, rank stabilization, custom modules
- **[Troubleshooting](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/peft/references/troubleshooting.md)** - Common errors, debugging, optimization
## Resources
- **GitHub**: https://github.com/huggingface/peft
- **Docs**: https://huggingface.co/docs/peft
- **LoRA Paper**: arXiv:2106.09685
- **QLoRA Paper**: arXiv:2305.14314
- **Models**: https://huggingface.co/models?library=peft

376
website/docs/user-guide/skills/optional/mlops/mlops-pinecone.md Normal file
View file

@ -0,0 +1,376 @@
---
title: "Pinecone — Managed vector database for production AI applications"
sidebar_label: "Pinecone"
description: "Managed vector database for production AI applications"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Pinecone
Managed vector database for production AI applications. Fully managed, auto-scaling, with hybrid search (dense + sparse), metadata filtering, and namespaces. Low latency (&lt;100ms p95). Use for production RAG, recommendation systems, or semantic search at scale. Best for serverless, managed infrastructure.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/pinecone` |
| Path | `optional-skills/mlops/pinecone` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `pinecone-client` |
| Tags | `RAG`, `Pinecone`, `Vector Database`, `Managed Service`, `Serverless`, `Hybrid Search`, `Production`, `Auto-Scaling`, `Low Latency`, `Recommendations` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Pinecone - Managed Vector Database
The vector database for production AI applications.
## When to use Pinecone
**Use when:**
- Need managed, serverless vector database
- Production RAG applications
- Auto-scaling required
- Low latency critical (&lt;100ms)
- Don't want to manage infrastructure
- Need hybrid search (dense + sparse vectors)
**Metrics**:
- Fully managed SaaS
- Auto-scales to billions of vectors
- **p95 latency &lt;100ms**
- 99.9% uptime SLA
**Use alternatives instead**:
- **Chroma**: Self-hosted, open-source
- **FAISS**: Offline, pure similarity search
- **Weaviate**: Self-hosted with more features
## Quick start
### Installation
```bash
pip install pinecone-client
```
### Basic usage
```python
from pinecone import Pinecone, ServerlessSpec
# Initialize
pc = Pinecone(api_key="your-api-key")
# Create index
pc.create_index(
name="my-index",
dimension=1536, # Must match embedding dimension
metric="cosine", # or "euclidean", "dotproduct"
spec=ServerlessSpec(cloud="aws", region="us-east-1")
)
# Connect to index
index = pc.Index("my-index")
# Upsert vectors
index.upsert(vectors=[
{"id": "vec1", "values": [0.1, 0.2, ...], "metadata": {"category": "A"}},
{"id": "vec2", "values": [0.3, 0.4, ...], "metadata": {"category": "B"}}
])
# Query
results = index.query(
vector=[0.1, 0.2, ...],
top_k=5,
include_metadata=True
)
print(results["matches"])
```
## Core operations
### Create index
```python
# Serverless (recommended)
pc.create_index(
name="my-index",
dimension=1536,
metric="cosine",
spec=ServerlessSpec(
cloud="aws", # or "gcp", "azure"
region="us-east-1"
)
)
# Pod-based (for consistent performance)
from pinecone import PodSpec
pc.create_index(
name="my-index",
dimension=1536,
metric="cosine",
spec=PodSpec(
environment="us-east1-gcp",
pod_type="p1.x1"
)
)
```
### Upsert vectors
```python
# Single upsert
index.upsert(vectors=[
{
"id": "doc1",
"values": [0.1, 0.2, ...], # 1536 dimensions
"metadata": {
"text": "Document content",
"category": "tutorial",
"timestamp": "2025-01-01"
}
}
])
# Batch upsert (recommended)
vectors = [
{"id": f"vec{i}", "values": embedding, "metadata": metadata}
for i, (embedding, metadata) in enumerate(zip(embeddings, metadatas))
]
index.upsert(vectors=vectors, batch_size=100)
```
### Query vectors
```python
# Basic query
results = index.query(
vector=[0.1, 0.2, ...],
top_k=10,
include_metadata=True,
include_values=False
)
# With metadata filtering
results = index.query(
vector=[0.1, 0.2, ...],
top_k=5,
filter={"category": {"$eq": "tutorial"}}
)
# Namespace query
results = index.query(
vector=[0.1, 0.2, ...],
top_k=5,
namespace="production"
)
# Access results
for match in results["matches"]:
print(f"ID: {match['id']}")
print(f"Score: {match['score']}")
print(f"Metadata: {match['metadata']}")
```
### Metadata filtering
```python
# Exact match
filter = {"category": "tutorial"}
# Comparison
filter = {"price": {"$gte": 100}} # $gt, $gte, $lt, $lte, $ne
# Logical operators
filter = {
"$and": [
{"category": "tutorial"},
{"difficulty": {"$lte": 3}}
]
} # Also: $or
# In operator
filter = {"tags": {"$in": ["python", "ml"]}}
```
## Namespaces
```python
# Partition data by namespace
index.upsert(
vectors=[{"id": "vec1", "values": [...]}],
namespace="user-123"
)
# Query specific namespace
results = index.query(
vector=[...],
namespace="user-123",
top_k=5
)
# List namespaces
stats = index.describe_index_stats()
print(stats['namespaces'])
```
## Hybrid search (dense + sparse)
```python
# Upsert with sparse vectors
index.upsert(vectors=[
{
"id": "doc1",
"values": [0.1, 0.2, ...], # Dense vector
"sparse_values": {
"indices": [10, 45, 123], # Token IDs
"values": [0.5, 0.3, 0.8] # TF-IDF scores
},
"metadata": {"text": "..."}
}
])
# Hybrid query
results = index.query(
vector=[0.1, 0.2, ...],
sparse_vector={
"indices": [10, 45],
"values": [0.5, 0.3]
},
top_k=5,
alpha=0.5 # 0=sparse, 1=dense, 0.5=hybrid
)
```
## LangChain integration
```python
from langchain_pinecone import PineconeVectorStore
from langchain_openai import OpenAIEmbeddings
# Create vector store
vectorstore = PineconeVectorStore.from_documents(
documents=docs,
embedding=OpenAIEmbeddings(),
index_name="my-index"
)
# Query
results = vectorstore.similarity_search("query", k=5)
# With metadata filter
results = vectorstore.similarity_search(
"query",
k=5,
filter={"category": "tutorial"}
)
# As retriever
retriever = vectorstore.as_retriever(search_kwargs={"k": 10})
```
## LlamaIndex integration
```python
from llama_index.vector_stores.pinecone import PineconeVectorStore
# Connect to Pinecone
pc = Pinecone(api_key="your-key")
pinecone_index = pc.Index("my-index")
# Create vector store
vector_store = PineconeVectorStore(pinecone_index=pinecone_index)
# Use in LlamaIndex
from llama_index.core import StorageContext, VectorStoreIndex
storage_context = StorageContext.from_defaults(vector_store=vector_store)
index = VectorStoreIndex.from_documents(documents, storage_context=storage_context)
```
## Index management
```python
# List indices
indexes = pc.list_indexes()
# Describe index
index_info = pc.describe_index("my-index")
print(index_info)
# Get index stats
stats = index.describe_index_stats()
print(f"Total vectors: {stats['total_vector_count']}")
print(f"Namespaces: {stats['namespaces']}")
# Delete index
pc.delete_index("my-index")
```
## Delete vectors
```python
# Delete by ID
index.delete(ids=["vec1", "vec2"])
# Delete by filter
index.delete(filter={"category": "old"})
# Delete all in namespace
index.delete(delete_all=True, namespace="test")
# Delete entire index
index.delete(delete_all=True)
```
## Best practices
1. **Use serverless** - Auto-scaling, cost-effective
2. **Batch upserts** - More efficient (100-200 per batch)
3. **Add metadata** - Enable filtering
4. **Use namespaces** - Isolate data by user/tenant
5. **Monitor usage** - Check Pinecone dashboard
6. **Optimize filters** - Index frequently filtered fields
7. **Test with free tier** - 1 index, 100K vectors free
8. **Use hybrid search** - Better quality
9. **Set appropriate dimensions** - Match embedding model
10. **Regular backups** - Export important data
## Performance
| Operation | Latency | Notes |
|-----------|---------|-------|
| Upsert | ~50-100ms | Per batch |
| Query (p50) | ~50ms | Depends on index size |
| Query (p95) | ~100ms | SLA target |
| Metadata filter | ~+10-20ms | Additional overhead |
## Pricing (as of 2025)
**Serverless**:
- $0.096 per million read units
- $0.06 per million write units
- $0.06 per GB storage/month
**Free tier**:
- 1 serverless index
- 100K vectors (1536 dimensions)
- Great for prototyping
## Resources
- **Website**: https://www.pinecone.io
- **Docs**: https://docs.pinecone.io
- **Console**: https://app.pinecone.io
- **Pricing**: https://www.pinecone.io/pricing

144
website/docs/user-guide/skills/optional/mlops/mlops-pytorch-fsdp.md Normal file
View file

File diff suppressed because one or more lines are too long

364
website/docs/user-guide/skills/optional/mlops/mlops-pytorch-lightning.md Normal file
View file

@ -0,0 +1,364 @@
---
title: "Pytorch Lightning"
sidebar_label: "Pytorch Lightning"
description: "High-level PyTorch framework with Trainer class, automatic distributed training (DDP/FSDP/DeepSpeed), callbacks system, and minimal boilerplate"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Pytorch Lightning
High-level PyTorch framework with Trainer class, automatic distributed training (DDP/FSDP/DeepSpeed), callbacks system, and minimal boilerplate. Scales from laptop to supercomputer with same code. Use when you want clean training loops with built-in best practices.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/pytorch-lightning` |
| Path | `optional-skills/mlops/pytorch-lightning` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `lightning`, `torch`, `transformers` |
| Tags | `PyTorch Lightning`, `Training Framework`, `Distributed Training`, `DDP`, `FSDP`, `DeepSpeed`, `High-Level API`, `Callbacks`, `Best Practices`, `Scalable` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# PyTorch Lightning - High-Level Training Framework
## Quick start
PyTorch Lightning organizes PyTorch code to eliminate boilerplate while maintaining flexibility.
**Installation**:
```bash
pip install lightning
```
**Convert PyTorch to Lightning** (3 steps):
```python
import lightning as L
import torch
from torch import nn
from torch.utils.data import DataLoader, Dataset
# Step 1: Define LightningModule (organize your PyTorch code)
class LitModel(L.LightningModule):
def __init__(self, hidden_size=128):
super().__init__()
self.model = nn.Sequential(
nn.Linear(28 * 28, hidden_size),
nn.ReLU(),
nn.Linear(hidden_size, 10)
)
def training_step(self, batch, batch_idx):
x, y = batch
y_hat = self.model(x)
loss = nn.functional.cross_entropy(y_hat, y)
self.log('train_loss', loss) # Auto-logged to TensorBoard
return loss
def configure_optimizers(self):
return torch.optim.Adam(self.parameters(), lr=1e-3)
# Step 2: Create data
train_loader = DataLoader(train_dataset, batch_size=32)
# Step 3: Train with Trainer (handles everything else!)
trainer = L.Trainer(max_epochs=10, accelerator='gpu', devices=2)
model = LitModel()
trainer.fit(model, train_loader)
```
**That's it!** Trainer handles:
- GPU/TPU/CPU switching
- Distributed training (DDP, FSDP, DeepSpeed)
- Mixed precision (FP16, BF16)
- Gradient accumulation
- Checkpointing
- Logging
- Progress bars
## Common workflows
### Workflow 1: From PyTorch to Lightning
**Original PyTorch code**:
```python
model = MyModel()
optimizer = torch.optim.Adam(model.parameters())
model.to('cuda')
for epoch in range(max_epochs):
for batch in train_loader:
batch = batch.to('cuda')
optimizer.zero_grad()
loss = model(batch)
loss.backward()
optimizer.step()
```
**Lightning version**:
```python
class LitModel(L.LightningModule):
def __init__(self):
super().__init__()
self.model = MyModel()
def training_step(self, batch, batch_idx):
loss = self.model(batch) # No .to('cuda') needed!
return loss
def configure_optimizers(self):
return torch.optim.Adam(self.parameters())
# Train
trainer = L.Trainer(max_epochs=10, accelerator='gpu')
trainer.fit(LitModel(), train_loader)
```
**Benefits**: 40+ lines → 15 lines, no device management, automatic distributed
### Workflow 2: Validation and testing
```python
class LitModel(L.LightningModule):
def __init__(self):
super().__init__()
self.model = MyModel()
def training_step(self, batch, batch_idx):
x, y = batch
y_hat = self.model(x)
loss = nn.functional.cross_entropy(y_hat, y)
self.log('train_loss', loss)
return loss
def validation_step(self, batch, batch_idx):
x, y = batch
y_hat = self.model(x)
val_loss = nn.functional.cross_entropy(y_hat, y)
acc = (y_hat.argmax(dim=1) == y).float().mean()
self.log('val_loss', val_loss)
self.log('val_acc', acc)
def test_step(self, batch, batch_idx):
x, y = batch
y_hat = self.model(x)
test_loss = nn.functional.cross_entropy(y_hat, y)
self.log('test_loss', test_loss)
def configure_optimizers(self):
return torch.optim.Adam(self.parameters(), lr=1e-3)
# Train with validation
trainer = L.Trainer(max_epochs=10)
trainer.fit(model, train_loader, val_loader)
# Test
trainer.test(model, test_loader)
```
**Automatic features**:
- Validation runs every epoch by default
- Metrics logged to TensorBoard
- Best model checkpointing based on val_loss
### Workflow 3: Distributed training (DDP)
```python
# Same code as single GPU!
model = LitModel()
# 8 GPUs with DDP (automatic!)
trainer = L.Trainer(
accelerator='gpu',
devices=8,
strategy='ddp' # Or 'fsdp', 'deepspeed'
)
trainer.fit(model, train_loader)
```
**Launch**:
```bash
# Single command, Lightning handles the rest
python train.py
```
**No changes needed**:
- Automatic data distribution
- Gradient synchronization
- Multi-node support (just set `num_nodes=2`)
### Workflow 4: Callbacks for monitoring
```python
from lightning.pytorch.callbacks import ModelCheckpoint, EarlyStopping, LearningRateMonitor
# Create callbacks
checkpoint = ModelCheckpoint(
monitor='val_loss',
mode='min',
save_top_k=3,
filename='model-{epoch:02d}-{val_loss:.2f}'
)
early_stop = EarlyStopping(
monitor='val_loss',
patience=5,
mode='min'
)
lr_monitor = LearningRateMonitor(logging_interval='epoch')
# Add to Trainer
trainer = L.Trainer(
max_epochs=100,
callbacks=[checkpoint, early_stop, lr_monitor]
)
trainer.fit(model, train_loader, val_loader)
```
**Result**:
- Auto-saves best 3 models
- Stops early if no improvement for 5 epochs
- Logs learning rate to TensorBoard
### Workflow 5: Learning rate scheduling
```python
class LitModel(L.LightningModule):
# ... (training_step, etc.)
def configure_optimizers(self):
optimizer = torch.optim.Adam(self.parameters(), lr=1e-3)
# Cosine annealing
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer,
T_max=100,
eta_min=1e-5
)
return {
'optimizer': optimizer,
'lr_scheduler': {
'scheduler': scheduler,
'interval': 'epoch', # Update per epoch
'frequency': 1
}
}
# Learning rate auto-logged!
trainer = L.Trainer(max_epochs=100)
trainer.fit(model, train_loader)
```
## When to use vs alternatives
**Use PyTorch Lightning when**:
- Want clean, organized code
- Need production-ready training loops
- Switching between single GPU, multi-GPU, TPU
- Want built-in callbacks and logging
- Team collaboration (standardized structure)
**Key advantages**:
- **Organized**: Separates research code from engineering
- **Automatic**: DDP, FSDP, DeepSpeed with 1 line
- **Callbacks**: Modular training extensions
- **Reproducible**: Less boilerplate = fewer bugs
- **Tested**: 1M+ downloads/month, battle-tested
**Use alternatives instead**:
- **Accelerate**: Minimal changes to existing code, more flexibility
- **Ray Train**: Multi-node orchestration, hyperparameter tuning
- **Raw PyTorch**: Maximum control, learning purposes
- **Keras**: TensorFlow ecosystem
## Common issues
**Issue: Loss not decreasing**
Check data and model setup:
```python
# Add to training_step
def training_step(self, batch, batch_idx):
if batch_idx == 0:
print(f"Batch shape: {batch[0].shape}")
print(f"Labels: {batch[1]}")
loss = ...
return loss
```
**Issue: Out of memory**
Reduce batch size or use gradient accumulation:
```python
trainer = L.Trainer(
accumulate_grad_batches=4, # Effective batch = batch_size × 4
precision='bf16' # Or 'fp16', reduces memory 50%
)
```
**Issue: Validation not running**
Ensure you pass val_loader:
```python
# WRONG
trainer.fit(model, train_loader)
# CORRECT
trainer.fit(model, train_loader, val_loader)
```
**Issue: DDP spawns multiple processes unexpectedly**
Lightning auto-detects GPUs. Explicitly set devices:
```python
# Test on CPU first
trainer = L.Trainer(accelerator='cpu', devices=1)
# Then GPU
trainer = L.Trainer(accelerator='gpu', devices=1)
```
## Advanced topics
**Callbacks**: See [references/callbacks.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/pytorch-lightning/references/callbacks.md) for EarlyStopping, ModelCheckpoint, custom callbacks, and callback hooks.
**Distributed strategies**: See [references/distributed.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/pytorch-lightning/references/distributed.md) for DDP, FSDP, DeepSpeed ZeRO integration, multi-node setup.
**Hyperparameter tuning**: See [references/hyperparameter-tuning.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/pytorch-lightning/references/hyperparameter-tuning.md) for integration with Optuna, Ray Tune, and WandB sweeps.
## Hardware requirements
- **CPU**: Works (good for debugging)
- **Single GPU**: Works
- **Multi-GPU**: DDP (default), FSDP, or DeepSpeed
- **Multi-node**: DDP, FSDP, DeepSpeed
- **TPU**: Supported (8 cores)
- **Apple MPS**: Supported
**Precision options**:
- FP32 (default)
- FP16 (V100, older GPUs)
- BF16 (A100/H100, recommended)
- FP8 (H100)
## Resources
- Docs: https://lightning.ai/docs/pytorch/stable/
- GitHub: https://github.com/Lightning-AI/pytorch-lightning ⭐ 29,000+
- Version: 2.5.5+
- Examples: https://github.com/Lightning-AI/pytorch-lightning/tree/master/examples
- Discord: https://discord.gg/lightning-ai
- Used by: Kaggle winners, research labs, production teams

513
website/docs/user-guide/skills/optional/mlops/mlops-qdrant.md Normal file
View file

@ -0,0 +1,513 @@
---
title: "Qdrant Vector Search — High-performance vector similarity search engine for RAG and semantic search"
sidebar_label: "Qdrant Vector Search"
description: "High-performance vector similarity search engine for RAG and semantic search"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Qdrant Vector Search
High-performance vector similarity search engine for RAG and semantic search. Use when building production RAG systems requiring fast nearest neighbor search, hybrid search with filtering, or scalable vector storage with Rust-powered performance.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/qdrant` |
| Path | `optional-skills/mlops/qdrant` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `qdrant-client>=1.12.0` |
| Tags | `RAG`, `Vector Search`, `Qdrant`, `Semantic Search`, `Embeddings`, `Similarity Search`, `HNSW`, `Production`, `Distributed` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Qdrant - Vector Similarity Search Engine
High-performance vector database written in Rust for production RAG and semantic search.
## When to use Qdrant
**Use Qdrant when:**
- Building production RAG systems requiring low latency
- Need hybrid search (vectors + metadata filtering)
- Require horizontal scaling with sharding/replication
- Want on-premise deployment with full data control
- Need multi-vector storage per record (dense + sparse)
- Building real-time recommendation systems
**Key features:**
- **Rust-powered**: Memory-safe, high performance
- **Rich filtering**: Filter by any payload field during search
- **Multiple vectors**: Dense, sparse, multi-dense per point
- **Quantization**: Scalar, product, binary for memory efficiency
- **Distributed**: Raft consensus, sharding, replication
- **REST + gRPC**: Both APIs with full feature parity
**Use alternatives instead:**
- **Chroma**: Simpler setup, embedded use cases
- **FAISS**: Maximum raw speed, research/batch processing
- **Pinecone**: Fully managed, zero ops preferred
- **Weaviate**: GraphQL preference, built-in vectorizers
## Quick start
### Installation
```bash
# Python client
pip install qdrant-client
# Docker (recommended for development)
docker run -p 6333:6333 -p 6334:6334 qdrant/qdrant
# Docker with persistent storage
docker run -p 6333:6333 -p 6334:6334 \
-v $(pwd)/qdrant_storage:/qdrant/storage \
qdrant/qdrant
```
### Basic usage
```python
from qdrant_client import QdrantClient
from qdrant_client.models import Distance, VectorParams, PointStruct
# Connect to Qdrant
client = QdrantClient(host="localhost", port=6333)
# Create collection
client.create_collection(
collection_name="documents",
vectors_config=VectorParams(size=384, distance=Distance.COSINE)
)
# Insert vectors with payload
client.upsert(
collection_name="documents",
points=[
PointStruct(
id=1,
vector=[0.1, 0.2, ...], # 384-dim vector
payload={"title": "Doc 1", "category": "tech"}
),
PointStruct(
id=2,
vector=[0.3, 0.4, ...],
payload={"title": "Doc 2", "category": "science"}
)
]
)
# Search with filtering
results = client.search(
collection_name="documents",
query_vector=[0.15, 0.25, ...],
query_filter={
"must": [{"key": "category", "match": {"value": "tech"}}]
},
limit=10
)
for point in results:
print(f"ID: {point.id}, Score: {point.score}, Payload: {point.payload}")
```
## Core concepts
### Points - Basic data unit
```python
from qdrant_client.models import PointStruct
# Point = ID + Vector(s) + Payload
point = PointStruct(
id=123, # Integer or UUID string
vector=[0.1, 0.2, 0.3, ...], # Dense vector
payload={ # Arbitrary JSON metadata
"title": "Document title",
"category": "tech",
"timestamp": 1699900000,
"tags": ["python", "ml"]
}
)
# Batch upsert (recommended)
client.upsert(
collection_name="documents",
points=[point1, point2, point3],
wait=True # Wait for indexing
)
```
### Collections - Vector containers
```python
from qdrant_client.models import VectorParams, Distance, HnswConfigDiff
# Create with HNSW configuration
client.create_collection(
collection_name="documents",
vectors_config=VectorParams(
size=384, # Vector dimensions
distance=Distance.COSINE # COSINE, EUCLID, DOT, MANHATTAN
),
hnsw_config=HnswConfigDiff(
m=16, # Connections per node (default 16)
ef_construct=100, # Build-time accuracy (default 100)
full_scan_threshold=10000 # Switch to brute force below this
),
on_disk_payload=True # Store payload on disk
)
# Collection info
info = client.get_collection("documents")
print(f"Points: {info.points_count}, Vectors: {info.vectors_count}")
```
### Distance metrics
| Metric | Use Case | Range |
|--------|----------|-------|
| `COSINE` | Text embeddings, normalized vectors | 0 to 2 |
| `EUCLID` | Spatial data, image features | 0 to ∞ |
| `DOT` | Recommendations, unnormalized | -∞ to ∞ |
| `MANHATTAN` | Sparse features, discrete data | 0 to ∞ |
## Search operations
### Basic search
```python
# Simple nearest neighbor search
results = client.search(
collection_name="documents",
query_vector=[0.1, 0.2, ...],
limit=10,
with_payload=True,
with_vectors=False # Don't return vectors (faster)
)
```
### Filtered search
```python
from qdrant_client.models import Filter, FieldCondition, MatchValue, Range
# Complex filtering
results = client.search(
collection_name="documents",
query_vector=query_embedding,
query_filter=Filter(
must=[
FieldCondition(key="category", match=MatchValue(value="tech")),
FieldCondition(key="timestamp", range=Range(gte=1699000000))
],
must_not=[
FieldCondition(key="status", match=MatchValue(value="archived"))
]
),
limit=10
)
# Shorthand filter syntax
results = client.search(
collection_name="documents",
query_vector=query_embedding,
query_filter={
"must": [
{"key": "category", "match": {"value": "tech"}},
{"key": "price", "range": {"gte": 10, "lte": 100}}
]
},
limit=10
)
```
### Batch search
```python
from qdrant_client.models import SearchRequest
# Multiple queries in one request
results = client.search_batch(
collection_name="documents",
requests=[
SearchRequest(vector=[0.1, ...], limit=5),
SearchRequest(vector=[0.2, ...], limit=5, filter={"must": [...]}),
SearchRequest(vector=[0.3, ...], limit=10)
]
)
```
## RAG integration
### With sentence-transformers
```python
from sentence_transformers import SentenceTransformer
from qdrant_client import QdrantClient
from qdrant_client.models import VectorParams, Distance, PointStruct
# Initialize
encoder = SentenceTransformer("all-MiniLM-L6-v2")
client = QdrantClient(host="localhost", port=6333)
# Create collection
client.create_collection(
collection_name="knowledge_base",
vectors_config=VectorParams(size=384, distance=Distance.COSINE)
)
# Index documents
documents = [
{"id": 1, "text": "Python is a programming language", "source": "wiki"},
{"id": 2, "text": "Machine learning uses algorithms", "source": "textbook"},
]
points = [
PointStruct(
id=doc["id"],
vector=encoder.encode(doc["text"]).tolist(),
payload={"text": doc["text"], "source": doc["source"]}
)
for doc in documents
]
client.upsert(collection_name="knowledge_base", points=points)
# RAG retrieval
def retrieve(query: str, top_k: int = 5) -> list[dict]:
query_vector = encoder.encode(query).tolist()
results = client.search(
collection_name="knowledge_base",
query_vector=query_vector,
limit=top_k
)
return [{"text": r.payload["text"], "score": r.score} for r in results]
# Use in RAG pipeline
context = retrieve("What is Python?")
prompt = f"Context: {context}\n\nQuestion: What is Python?"
```
### With LangChain
```python
from langchain_community.vectorstores import Qdrant
from langchain_community.embeddings import HuggingFaceEmbeddings
embeddings = HuggingFaceEmbeddings(model_name="all-MiniLM-L6-v2")
vectorstore = Qdrant.from_documents(documents, embeddings, url="http://localhost:6333", collection_name="docs")
retriever = vectorstore.as_retriever(search_kwargs={"k": 5})
```
### With LlamaIndex
```python
from llama_index.vector_stores.qdrant import QdrantVectorStore
from llama_index.core import VectorStoreIndex, StorageContext
vector_store = QdrantVectorStore(client=client, collection_name="llama_docs")
storage_context = StorageContext.from_defaults(vector_store=vector_store)
index = VectorStoreIndex.from_documents(documents, storage_context=storage_context)
query_engine = index.as_query_engine()
```
## Multi-vector support
### Named vectors (different embedding models)
```python
from qdrant_client.models import VectorParams, Distance
# Collection with multiple vector types
client.create_collection(
collection_name="hybrid_search",
vectors_config={
"dense": VectorParams(size=384, distance=Distance.COSINE),
"sparse": VectorParams(size=30000, distance=Distance.DOT)
}
)
# Insert with named vectors
client.upsert(
collection_name="hybrid_search",
points=[
PointStruct(
id=1,
vector={
"dense": dense_embedding,
"sparse": sparse_embedding
},
payload={"text": "document text"}
)
]
)
# Search specific vector
results = client.search(
collection_name="hybrid_search",
query_vector=("dense", query_dense), # Specify which vector
limit=10
)
```
### Sparse vectors (BM25, SPLADE)
```python
from qdrant_client.models import SparseVectorParams, SparseIndexParams, SparseVector
# Collection with sparse vectors
client.create_collection(
collection_name="sparse_search",
vectors_config={},
sparse_vectors_config={"text": SparseVectorParams(index=SparseIndexParams(on_disk=False))}
)
# Insert sparse vector
client.upsert(
collection_name="sparse_search",
points=[PointStruct(id=1, vector={"text": SparseVector(indices=[1, 5, 100], values=[0.5, 0.8, 0.2])}, payload={"text": "document"})]
)
```
## Quantization (memory optimization)
```python
from qdrant_client.models import ScalarQuantization, ScalarQuantizationConfig, ScalarType
# Scalar quantization (4x memory reduction)
client.create_collection(
collection_name="quantized",
vectors_config=VectorParams(size=384, distance=Distance.COSINE),
quantization_config=ScalarQuantization(
scalar=ScalarQuantizationConfig(
type=ScalarType.INT8,
quantile=0.99, # Clip outliers
always_ram=True # Keep quantized in RAM
)
)
)
# Search with rescoring
results = client.search(
collection_name="quantized",
query_vector=query,
search_params={"quantization": {"rescore": True}}, # Rescore top results
limit=10
)
```
## Payload indexing
```python
from qdrant_client.models import PayloadSchemaType
# Create payload index for faster filtering
client.create_payload_index(
collection_name="documents",
field_name="category",
field_schema=PayloadSchemaType.KEYWORD
)
client.create_payload_index(
collection_name="documents",
field_name="timestamp",
field_schema=PayloadSchemaType.INTEGER
)
# Index types: KEYWORD, INTEGER, FLOAT, GEO, TEXT (full-text), BOOL
```
## Production deployment
### Qdrant Cloud
```python
from qdrant_client import QdrantClient
# Connect to Qdrant Cloud
client = QdrantClient(
url="https://your-cluster.cloud.qdrant.io",
api_key="your-api-key"
)
```
### Performance tuning
```python
# Optimize for search speed (higher recall)
client.update_collection(
collection_name="documents",
hnsw_config=HnswConfigDiff(ef_construct=200, m=32)
)
# Optimize for indexing speed (bulk loads)
client.update_collection(
collection_name="documents",
optimizer_config={"indexing_threshold": 20000}
)
```
## Best practices
1. **Batch operations** - Use batch upsert/search for efficiency
2. **Payload indexing** - Index fields used in filters
3. **Quantization** - Enable for large collections (>1M vectors)
4. **Sharding** - Use for collections >10M vectors
5. **On-disk storage** - Enable `on_disk_payload` for large payloads
6. **Connection pooling** - Reuse client instances
## Common issues
**Slow search with filters:**
```python
# Create payload index for filtered fields
client.create_payload_index(
collection_name="docs",
field_name="category",
field_schema=PayloadSchemaType.KEYWORD
)
```
**Out of memory:**
```python
# Enable quantization and on-disk storage
client.create_collection(
collection_name="large_collection",
vectors_config=VectorParams(size=384, distance=Distance.COSINE),
quantization_config=ScalarQuantization(...),
on_disk_payload=True
)
```
**Connection issues:**
```python
# Use timeout and retry
client = QdrantClient(
host="localhost",
port=6333,
timeout=30,
prefer_grpc=True # gRPC for better performance
)
```
## References
- **[Advanced Usage](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/qdrant/references/advanced-usage.md)** - Distributed mode, hybrid search, recommendations
- **[Troubleshooting](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/qdrant/references/troubleshooting.md)** - Common issues, debugging, performance tuning
## Resources
- **GitHub**: https://github.com/qdrant/qdrant (22k+ stars)
- **Docs**: https://qdrant.tech/documentation/
- **Python Client**: https://github.com/qdrant/qdrant-client
- **Cloud**: https://cloud.qdrant.io
- **Version**: 1.12.0+
- **License**: Apache 2.0

406
website/docs/user-guide/skills/optional/mlops/mlops-saelens.md Normal file
View file

@ -0,0 +1,406 @@
---
title: "Sparse Autoencoder Training"
sidebar_label: "Sparse Autoencoder Training"
description: "Provides guidance for training and analyzing Sparse Autoencoders (SAEs) using SAELens to decompose neural network activations into interpretable features"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Sparse Autoencoder Training
Provides guidance for training and analyzing Sparse Autoencoders (SAEs) using SAELens to decompose neural network activations into interpretable features. Use when discovering interpretable features, analyzing superposition, or studying monosemantic representations in language models.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/saelens` |
| Path | `optional-skills/mlops/saelens` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `sae-lens>=6.0.0`, `transformer-lens>=2.0.0`, `torch>=2.0.0` |
| Tags | `Sparse Autoencoders`, `SAE`, `Mechanistic Interpretability`, `Feature Discovery`, `Superposition` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# SAELens: Sparse Autoencoders for Mechanistic Interpretability
SAELens is the primary library for training and analyzing Sparse Autoencoders (SAEs) - a technique for decomposing polysemantic neural network activations into sparse, interpretable features. Based on Anthropic's groundbreaking research on monosemanticity.
**GitHub**: [jbloomAus/SAELens](https://github.com/jbloomAus/SAELens) (1,100+ stars)
## The Problem: Polysemanticity & Superposition
Individual neurons in neural networks are **polysemantic** - they activate in multiple, semantically distinct contexts. This happens because models use **superposition** to represent more features than they have neurons, making interpretability difficult.
**SAEs solve this** by decomposing dense activations into sparse, monosemantic features - typically only a small number of features activate for any given input, and each feature corresponds to an interpretable concept.
## When to Use SAELens
**Use SAELens when you need to:**
- Discover interpretable features in model activations
- Understand what concepts a model has learned
- Study superposition and feature geometry
- Perform feature-based steering or ablation
- Analyze safety-relevant features (deception, bias, harmful content)
**Consider alternatives when:**
- You need basic activation analysis → Use **TransformerLens** directly
- You want causal intervention experiments → Use **pyvene** or **TransformerLens**
- You need production steering → Consider direct activation engineering
## Installation
```bash
pip install sae-lens
```
Requirements: Python 3.10+, transformer-lens>=2.0.0
## Core Concepts
### What SAEs Learn
SAEs are trained to reconstruct model activations through a sparse bottleneck:
```
Input Activation → Encoder → Sparse Features → Decoder → Reconstructed Activation
(d_model) ↓ (d_sae >> d_model) ↓ (d_model)
sparsity reconstruction
penalty loss
```
**Loss Function**: `MSE(original, reconstructed) + L1_coefficient × L1(features)`
### Key Validation (Anthropic Research)
In "Towards Monosemanticity", human evaluators found **70% of SAE features genuinely interpretable**. Features discovered include:
- DNA sequences, legal language, HTTP requests
- Hebrew text, nutrition statements, code syntax
- Sentiment, named entities, grammatical structures
## Workflow 1: Loading and Analyzing Pre-trained SAEs
### Step-by-Step
```python
from transformer_lens import HookedTransformer
from sae_lens import SAE
# 1. Load model and pre-trained SAE
model = HookedTransformer.from_pretrained("gpt2-small", device="cuda")
sae, cfg_dict, sparsity = SAE.from_pretrained(
release="gpt2-small-res-jb",
sae_id="blocks.8.hook_resid_pre",
device="cuda"
)
# 2. Get model activations
tokens = model.to_tokens("The capital of France is Paris")
_, cache = model.run_with_cache(tokens)
activations = cache["resid_pre", 8] # [batch, pos, d_model]
# 3. Encode to SAE features
sae_features = sae.encode(activations) # [batch, pos, d_sae]
print(f"Active features: {(sae_features > 0).sum()}")
# 4. Find top features for each position
for pos in range(tokens.shape[1]):
top_features = sae_features[0, pos].topk(5)
token = model.to_str_tokens(tokens[0, pos:pos+1])[0]
print(f"Token '{token}': features {top_features.indices.tolist()}")
# 5. Reconstruct activations
reconstructed = sae.decode(sae_features)
reconstruction_error = (activations - reconstructed).norm()
```
### Available Pre-trained SAEs
| Release | Model | Layers |
|---------|-------|--------|
| `gpt2-small-res-jb` | GPT-2 Small | Multiple residual streams |
| `gemma-2b-res` | Gemma 2B | Residual streams |
| Various on HuggingFace | Search tag `saelens` | Various |
### Checklist
- [ ] Load model with TransformerLens
- [ ] Load matching SAE for target layer
- [ ] Encode activations to sparse features
- [ ] Identify top-activating features per token
- [ ] Validate reconstruction quality
## Workflow 2: Training a Custom SAE
### Step-by-Step
```python
from sae_lens import SAE, LanguageModelSAERunnerConfig, SAETrainingRunner
# 1. Configure training
cfg = LanguageModelSAERunnerConfig(
# Model
model_name="gpt2-small",
hook_name="blocks.8.hook_resid_pre",
hook_layer=8,
d_in=768, # Model dimension
# SAE architecture
architecture="standard", # or "gated", "topk"
d_sae=768 * 8, # Expansion factor of 8
activation_fn="relu",
# Training
lr=4e-4,
l1_coefficient=8e-5, # Sparsity penalty
l1_warm_up_steps=1000,
train_batch_size_tokens=4096,
training_tokens=100_000_000,
# Data
dataset_path="monology/pile-uncopyrighted",
context_size=128,
# Logging
log_to_wandb=True,
wandb_project="sae-training",
# Checkpointing
checkpoint_path="checkpoints",
n_checkpoints=5,
)
# 2. Train
trainer = SAETrainingRunner(cfg)
sae = trainer.run()
# 3. Evaluate
print(f"L0 (avg active features): {trainer.metrics['l0']}")
print(f"CE Loss Recovered: {trainer.metrics['ce_loss_score']}")
```
### Key Hyperparameters
| Parameter | Typical Value | Effect |
|-----------|---------------|--------|
| `d_sae` | 4-16× d_model | More features, higher capacity |
| `l1_coefficient` | 5e-5 to 1e-4 | Higher = sparser, less accurate |
| `lr` | 1e-4 to 1e-3 | Standard optimizer LR |
| `l1_warm_up_steps` | 500-2000 | Prevents early feature death |
### Evaluation Metrics
| Metric | Target | Meaning |
|--------|--------|---------|
| **L0** | 50-200 | Average active features per token |
| **CE Loss Score** | 80-95% | Cross-entropy recovered vs original |
| **Dead Features** | &lt;5% | Features that never activate |
| **Explained Variance** | >90% | Reconstruction quality |
### Checklist
- [ ] Choose target layer and hook point
- [ ] Set expansion factor (d_sae = 4-16× d_model)
- [ ] Tune L1 coefficient for desired sparsity
- [ ] Enable L1 warm-up to prevent dead features
- [ ] Monitor metrics during training (W&B)
- [ ] Validate L0 and CE loss recovery
- [ ] Check dead feature ratio
## Workflow 3: Feature Analysis and Steering
### Analyzing Individual Features
```python
from transformer_lens import HookedTransformer
from sae_lens import SAE
import torch
model = HookedTransformer.from_pretrained("gpt2-small", device="cuda")
sae, _, _ = SAE.from_pretrained(
release="gpt2-small-res-jb",
sae_id="blocks.8.hook_resid_pre",
device="cuda"
)
# Find what activates a specific feature
feature_idx = 1234
test_texts = [
"The scientist conducted an experiment",
"I love chocolate cake",
"The code compiles successfully",
"Paris is beautiful in spring",
]
for text in test_texts:
tokens = model.to_tokens(text)
_, cache = model.run_with_cache(tokens)
features = sae.encode(cache["resid_pre", 8])
activation = features[0, :, feature_idx].max().item()
print(f"{activation:.3f}: {text}")
```
### Feature Steering
```python
def steer_with_feature(model, sae, prompt, feature_idx, strength=5.0):
"""Add SAE feature direction to residual stream."""
tokens = model.to_tokens(prompt)
# Get feature direction from decoder
feature_direction = sae.W_dec[feature_idx] # [d_model]
def steering_hook(activation, hook):
# Add scaled feature direction at all positions
activation += strength * feature_direction
return activation
# Generate with steering
output = model.generate(
tokens,
max_new_tokens=50,
fwd_hooks=[("blocks.8.hook_resid_pre", steering_hook)]
)
return model.to_string(output[0])
```
### Feature Attribution
```python
# Which features most affect a specific output?
tokens = model.to_tokens("The capital of France is")
_, cache = model.run_with_cache(tokens)
# Get features at final position
features = sae.encode(cache["resid_pre", 8])[0, -1] # [d_sae]
# Get logit attribution per feature
# Feature contribution = feature_activation × decoder_weight × unembedding
W_dec = sae.W_dec # [d_sae, d_model]
W_U = model.W_U # [d_model, vocab]
# Contribution to "Paris" logit
paris_token = model.to_single_token(" Paris")
feature_contributions = features * (W_dec @ W_U[:, paris_token])
top_features = feature_contributions.topk(10)
print("Top features for 'Paris' prediction:")
for idx, val in zip(top_features.indices, top_features.values):
print(f" Feature {idx.item()}: {val.item():.3f}")
```
## Common Issues & Solutions
### Issue: High dead feature ratio
```python
# WRONG: No warm-up, features die early
cfg = LanguageModelSAERunnerConfig(
l1_coefficient=1e-4,
l1_warm_up_steps=0, # Bad!
)
# RIGHT: Warm-up L1 penalty
cfg = LanguageModelSAERunnerConfig(
l1_coefficient=8e-5,
l1_warm_up_steps=1000, # Gradually increase
use_ghost_grads=True, # Revive dead features
)
```
### Issue: Poor reconstruction (low CE recovery)
```python
# Reduce sparsity penalty
cfg = LanguageModelSAERunnerConfig(
l1_coefficient=5e-5, # Lower = better reconstruction
d_sae=768 * 16, # More capacity
)
```
### Issue: Features not interpretable
```python
# Increase sparsity (higher L1)
cfg = LanguageModelSAERunnerConfig(
l1_coefficient=1e-4, # Higher = sparser, more interpretable
)
# Or use TopK architecture
cfg = LanguageModelSAERunnerConfig(
architecture="topk",
activation_fn_kwargs={"k": 50}, # Exactly 50 active features
)
```
### Issue: Memory errors during training
```python
cfg = LanguageModelSAERunnerConfig(
train_batch_size_tokens=2048, # Reduce batch size
store_batch_size_prompts=4, # Fewer prompts in buffer
n_batches_in_buffer=8, # Smaller activation buffer
)
```
## Integration with Neuronpedia
Browse pre-trained SAE features at [neuronpedia.org](https://neuronpedia.org):
```python
# Features are indexed by SAE ID
# Example: gpt2-small layer 8 feature 1234
# → neuronpedia.org/gpt2-small/8-res-jb/1234
```
## Key Classes Reference
| Class | Purpose |
|-------|---------|
| `SAE` | Sparse Autoencoder model |
| `LanguageModelSAERunnerConfig` | Training configuration |
| `SAETrainingRunner` | Training loop manager |
| `ActivationsStore` | Activation collection and batching |
| `HookedSAETransformer` | TransformerLens + SAE integration |
## Reference Documentation
For detailed API documentation, tutorials, and advanced usage, see the `references/` folder:
| File | Contents |
|------|----------|
| [references/README.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/saelens/references/README.md) | Overview and quick start guide |
| [references/api.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/saelens/references/api.md) | Complete API reference for SAE, TrainingSAE, configurations |
| [references/tutorials.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/saelens/references/tutorials.md) | Step-by-step tutorials for training, analysis, steering |
## External Resources
### Tutorials
- [Basic Loading & Analysis](https://github.com/jbloomAus/SAELens/blob/main/tutorials/basic_loading_and_analysing.ipynb)
- [Training a Sparse Autoencoder](https://github.com/jbloomAus/SAELens/blob/main/tutorials/training_a_sparse_autoencoder.ipynb)
- [ARENA SAE Curriculum](https://www.lesswrong.com/posts/LnHowHgmrMbWtpkxx/intro-to-superposition-and-sparse-autoencoders-colab)
### Papers
- [Towards Monosemanticity](https://transformer-circuits.pub/2023/monosemantic-features) - Anthropic (2023)
- [Scaling Monosemanticity](https://transformer-circuits.pub/2024/scaling-monosemanticity/) - Anthropic (2024)
- [Sparse Autoencoders Find Highly Interpretable Features](https://arxiv.org/abs/2309.08600) - Cunningham et al. (ICLR 2024)
### Official Documentation
- [SAELens Docs](https://jbloomaus.github.io/SAELens/)
- [Neuronpedia](https://neuronpedia.org) - Feature browser
## SAE Architectures
| Architecture | Description | Use Case |
|--------------|-------------|----------|
| **Standard** | ReLU + L1 penalty | General purpose |
| **Gated** | Learned gating mechanism | Better sparsity control |
| **TopK** | Exactly K active features | Consistent sparsity |
```python
# TopK SAE (exactly 50 features active)
cfg = LanguageModelSAERunnerConfig(
architecture="topk",
activation_fn="topk",
activation_fn_kwargs={"k": 50},
)
```

236
website/docs/user-guide/skills/optional/mlops/mlops-simpo.md Normal file
View file

@ -0,0 +1,236 @@
---
title: "Simpo Training — Simple Preference Optimization for LLM alignment"
sidebar_label: "Simpo Training"
description: "Simple Preference Optimization for LLM alignment"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Simpo Training
Simple Preference Optimization for LLM alignment. Reference-free alternative to DPO with better performance (+6.4 points on AlpacaEval 2.0). No reference model needed, more efficient than DPO. Use for preference alignment when want simpler, faster training than DPO/PPO.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/simpo` |
| Path | `optional-skills/mlops/simpo` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `torch`, `transformers`, `datasets`, `trl`, `accelerate` |
| Tags | `Post-Training`, `SimPO`, `Preference Optimization`, `Alignment`, `DPO Alternative`, `Reference-Free`, `LLM Alignment`, `Efficient Training` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# SimPO - Simple Preference Optimization
## Quick start
SimPO is a reference-free preference optimization method that outperforms DPO without needing a reference model.
**Installation**:
```bash
# Create environment
conda create -n simpo python=3.10 && conda activate simpo
# Install PyTorch 2.2.2
# Visit: https://pytorch.org/get-started/locally/
# Install alignment-handbook
git clone https://github.com/huggingface/alignment-handbook.git
cd alignment-handbook
python -m pip install .
# Install Flash Attention 2
python -m pip install flash-attn --no-build-isolation
```
**Training** (Mistral 7B):
```bash
ACCELERATE_LOG_LEVEL=info accelerate launch \
--config_file accelerate_configs/deepspeed_zero3.yaml \
scripts/run_simpo.py \
training_configs/mistral-7b-base-simpo.yaml
```
## Common workflows
### Workflow 1: Train from base model (Mistral 7B)
**Config** (`mistral-7b-base-simpo.yaml`):
```yaml
# Model
model_name_or_path: mistralai/Mistral-7B-v0.1
torch_dtype: bfloat16
# Dataset
dataset_mixer:
HuggingFaceH4/ultrafeedback_binarized: 1.0
dataset_splits:
- train_prefs
- test_prefs
# SimPO hyperparameters
beta: 2.0 # Reward scaling (2.0-10.0)
gamma_beta_ratio: 0.5 # Target margin (0-1)
loss_type: sigmoid # sigmoid or hinge
sft_weight: 0.0 # Optional SFT regularization
# Training
learning_rate: 5e-7 # Critical: 3e-7 to 1e-6
num_train_epochs: 1
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
# Output
output_dir: ./outputs/mistral-7b-simpo
```
**Launch training**:
```bash
accelerate launch --config_file accelerate_configs/deepspeed_zero3.yaml \
scripts/run_simpo.py training_configs/mistral-7b-base-simpo.yaml
```
### Workflow 2: Fine-tune instruct model (Llama 3 8B)
**Config** (`llama3-8b-instruct-simpo.yaml`):
```yaml
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
dataset_mixer:
argilla/ultrafeedback-binarized-preferences-cleaned: 1.0
beta: 2.5
gamma_beta_ratio: 0.5
learning_rate: 5e-7
sft_weight: 0.1 # Add SFT loss to preserve capabilities
num_train_epochs: 1
per_device_train_batch_size: 2
gradient_accumulation_steps: 4
output_dir: ./outputs/llama3-8b-simpo
```
**Launch**:
```bash
accelerate launch --config_file accelerate_configs/deepspeed_zero3.yaml \
scripts/run_simpo.py training_configs/llama3-8b-instruct-simpo.yaml
```
### Workflow 3: Reasoning-intensive tasks (lower LR)
**For math/code tasks**:
```yaml
model_name_or_path: deepseek-ai/deepseek-math-7b-base
dataset_mixer:
argilla/distilabel-math-preference-dpo: 1.0
beta: 5.0 # Higher for stronger signal
gamma_beta_ratio: 0.7 # Larger margin
learning_rate: 3e-7 # Lower LR for reasoning
sft_weight: 0.0
num_train_epochs: 1
per_device_train_batch_size: 1
gradient_accumulation_steps: 16
```
## When to use vs alternatives
**Use SimPO when**:
- Want simpler training than DPO (no reference model)
- Have preference data (chosen/rejected pairs)
- Need better performance than DPO
- Limited compute resources
- Single-node training sufficient
**Algorithm selection**:
- **SimPO**: Simplest, best performance, no reference model
- **DPO**: Need reference model baseline, more conservative
- **PPO**: Maximum control, need reward model, complex setup
- **GRPO**: Memory-efficient RL, no critic
**Use alternatives instead**:
- **OpenRLHF**: Multi-node distributed training, PPO/GRPO
- **TRL**: Need multiple methods in one framework
- **DPO**: Established baseline comparison
## Common issues
**Issue: Loss divergence**
Reduce learning rate:
```yaml
learning_rate: 3e-7 # Reduce from 5e-7
```
Reduce beta:
```yaml
beta: 1.0 # Reduce from 2.0
```
**Issue: Model forgets capabilities**
Add SFT regularization:
```yaml
sft_weight: 0.1 # Add SFT loss component
```
**Issue: Poor preference separation**
Increase beta and margin:
```yaml
beta: 5.0 # Increase from 2.0
gamma_beta_ratio: 0.8 # Increase from 0.5
```
**Issue: OOM during training**
Reduce batch size:
```yaml
per_device_train_batch_size: 1
gradient_accumulation_steps: 16 # Maintain effective batch
```
Enable gradient checkpointing:
```yaml
gradient_checkpointing: true
```
## Advanced topics
**Loss functions**: See [references/loss-functions.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/simpo/references/loss-functions.md) for sigmoid vs hinge loss, mathematical formulations, and when to use each.
**Hyperparameter tuning**: See [references/hyperparameters.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/simpo/references/hyperparameters.md) for beta, gamma, learning rate selection guide, and model-size-specific recommendations.
**Dataset preparation**: See [references/datasets.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/simpo/references/datasets.md) for preference data formats, quality filtering, and custom dataset creation.
## Hardware requirements
- **GPU**: NVIDIA A100/H100 recommended
- **VRAM**:
- 7B model: 1× A100 40GB (DeepSpeed ZeRO-3)
- 8B model: 2× A100 40GB
- 70B model: 8× A100 80GB
- **Single-node**: DeepSpeed ZeRO-3 sufficient
- **Mixed precision**: BF16 recommended
**Memory optimization**:
- DeepSpeed ZeRO-3 (default config)
- Gradient checkpointing
- Flash Attention 2
## Resources
- Paper: https://arxiv.org/abs/2405.14734 (NeurIPS 2024)
- GitHub: https://github.com/princeton-nlp/SimPO
- Models: https://huggingface.co/princeton-nlp
- Alignment Handbook: https://github.com/huggingface/alignment-handbook

483
website/docs/user-guide/skills/optional/mlops/mlops-slime.md Normal file
View file

@ -0,0 +1,483 @@
---
title: "Slime Rl Training — Provides guidance for LLM post-training with RL using slime, a Megatron+SGLang framework"
sidebar_label: "Slime Rl Training"
description: "Provides guidance for LLM post-training with RL using slime, a Megatron+SGLang framework"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Slime Rl Training
Provides guidance for LLM post-training with RL using slime, a Megatron+SGLang framework. Use when training GLM models, implementing custom data generation workflows, or needing tight Megatron-LM integration for RL scaling.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/slime` |
| Path | `optional-skills/mlops/slime` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `sglang-router>=0.2.3`, `ray`, `torch>=2.0.0`, `transformers>=4.40.0` |
| Tags | `Reinforcement Learning`, `Megatron-LM`, `SGLang`, `GRPO`, `Post-Training`, `GLM` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# slime: LLM Post-Training Framework for RL Scaling
slime is an LLM post-training framework from Tsinghua's THUDM team, powering GLM-4.5, GLM-4.6, and GLM-4.7. It connects Megatron-LM for training with SGLang for high-throughput rollout generation.
## When to Use slime
**Choose slime when you need:**
- Megatron-LM native training with SGLang inference
- Custom data generation workflows with flexible data buffers
- Training GLM, Qwen3, DeepSeek V3, or Llama 3 models
- Research-grade framework with production backing (Z.ai)
**Consider alternatives when:**
- You need enterprise-grade stability features → use **miles**
- You want flexible backend swapping → use **verl**
- You need PyTorch-native abstractions → use **torchforge**
## Key Features
- **Training**: Megatron-LM with full parallelism support (TP, PP, DP, SP)
- **Rollout**: SGLang-based high-throughput generation with router
- **Data Buffer**: Flexible prompt management and sample storage
- **Models**: GLM-4.x, Qwen3, DeepSeek V3/R1, Llama 3
## Architecture Overview
```
┌─────────────────────────────────────────────────────────┐
│ Data Buffer │
│ - Prompt initialization and management │
│ - Custom data generation and filtering │
│ - Rollout sample storage │
└─────────────┬───────────────────────────┬───────────────┘
│ │
┌─────────────▼───────────┐ ┌─────────────▼───────────────┐
│ Training (Megatron-LM) │ │ Rollout (SGLang + Router) │
│ - Actor model training │ │ - Response generation │
│ - Critic (optional) │ │ - Reward/verifier output │
│ - Weight sync to rollout│ │ - Multi-turn support │
└─────────────────────────┘ └─────────────────────────────┘
```
## Installation
```bash
# Recommended: Docker
docker pull slimerl/slime:latest
docker run --rm --gpus all --ipc=host --shm-size=16g \
-it slimerl/slime:latest /bin/bash
# Inside container
cd /root/slime && pip install -e . --no-deps
```
### From Source
```bash
git clone https://github.com/THUDM/slime.git
cd slime
pip install -r requirements.txt
pip install -e .
```
## Quick Start: GRPO Training
```bash
# Source model configuration
source scripts/models/qwen3-4B.sh
# Launch training
python train.py \
--actor-num-nodes 1 \
--actor-num-gpus-per-node 4 \
--rollout-num-gpus 4 \
--advantage-estimator grpo \
--use-kl-loss --kl-loss-coef 0.001 \
--rollout-batch-size 32 \
--n-samples-per-prompt 8 \
--global-batch-size 256 \
--num-rollout 3000 \
--prompt-data /path/to/data.jsonl \
${MODEL_ARGS[@]} ${CKPT_ARGS[@]}
```
---
## Workflow 1: Standard GRPO Training
Use this workflow for training reasoning models with group-relative advantages.
### Prerequisites Checklist
- [ ] Docker environment or Megatron-LM + SGLang installed
- [ ] Model checkpoint (HuggingFace or Megatron format)
- [ ] Training data in JSONL format
### Step 1: Prepare Data
```python
# data.jsonl format
{"prompt": "What is 2 + 2?", "label": "4"}
{"prompt": "Solve: 3x = 12", "label": "x = 4"}
```
Or with chat format:
```python
{
"prompt": [
{"role": "system", "content": "You are a math tutor."},
{"role": "user", "content": "What is 15 + 27?"}
],
"label": "42"
}
```
### Step 2: Configure Model
Choose a pre-configured model script:
```bash
# List available models
ls scripts/models/
# glm4-9B.sh, qwen3-4B.sh, qwen3-30B-A3B.sh, deepseek-v3.sh, llama3-8B.sh, ...
# Source your model
source scripts/models/qwen3-4B.sh
```
### Step 3: Launch Training
```bash
python train.py \
--actor-num-nodes 1 \
--actor-num-gpus-per-node 8 \
--rollout-num-gpus 8 \
--advantage-estimator grpo \
--use-kl-loss \
--kl-loss-coef 0.001 \
--prompt-data /path/to/train.jsonl \
--input-key prompt \
--label-key label \
--apply-chat-template \
--rollout-batch-size 32 \
--n-samples-per-prompt 8 \
--global-batch-size 256 \
--num-rollout 3000 \
--save-interval 100 \
--eval-interval 50 \
${MODEL_ARGS[@]}
```
### Step 4: Monitor Training
- [ ] Check TensorBoard: `tensorboard --logdir outputs/`
- [ ] Verify reward curves are increasing
- [ ] Monitor GPU utilization across nodes
---
## Workflow 2: Asynchronous Training
Use async mode for higher throughput by overlapping rollout and training.
### When to Use Async
- Large models with long generation times
- High GPU idle time in synchronous mode
- Sufficient memory for buffering
### Launch Async Training
```bash
python train_async.py \
--actor-num-nodes 1 \
--actor-num-gpus-per-node 8 \
--rollout-num-gpus 8 \
--advantage-estimator grpo \
--async-buffer-size 4 \
--prompt-data /path/to/train.jsonl \
${MODEL_ARGS[@]}
```
### Async-Specific Parameters
```bash
--async-buffer-size 4 # Number of rollouts to buffer
--update-weights-interval 2 # Sync weights every N rollouts
```
---
## Workflow 3: Multi-Turn Agentic Training
Use this workflow for training agents with tool use or multi-step reasoning.
### Prerequisites
- [ ] Custom generate function for multi-turn logic
- [ ] Tool/environment interface
### Step 1: Define Custom Generate Function
```python
# custom_generate.py
async def custom_generate(args, samples, evaluation=False):
"""Multi-turn generation with tool calling."""
for sample in samples:
conversation = sample.prompt
for turn in range(args.max_turns):
# Generate response
response = await generate_single(conversation)
# Check for tool call
tool_call = extract_tool_call(response)
if tool_call:
tool_result = execute_tool(tool_call)
conversation.append({"role": "assistant", "content": response})
conversation.append({"role": "tool", "content": tool_result})
else:
break
sample.response = response
sample.reward = compute_reward(sample)
return samples
```
### Step 2: Launch with Custom Function
```bash
python train.py \
--custom-generate-function-path custom_generate.py \
--max-turns 5 \
--prompt-data /path/to/agent_data.jsonl \
${MODEL_ARGS[@]}
```
See `examples/search-r1/` for a complete multi-turn search example.
---
## Configuration Reference
### Three Argument Categories
slime uses three types of arguments:
**1. Megatron Arguments** (passed directly):
```bash
--tensor-model-parallel-size 2
--pipeline-model-parallel-size 1
--num-layers 32
--hidden-size 4096
```
**2. SGLang Arguments** (prefixed with `--sglang-`):
```bash
--sglang-mem-fraction-static 0.8
--sglang-context-length 8192
--sglang-log-level INFO
```
**3. slime Arguments**:
```bash
# Resource allocation
--actor-num-nodes 1
--actor-num-gpus-per-node 8
--rollout-num-gpus 8
--colocate # Share GPUs between training/inference
# Data
--prompt-data /path/to/data.jsonl
--input-key prompt
--label-key label
# Training loop
--num-rollout 3000
--rollout-batch-size 32
--n-samples-per-prompt 8
--global-batch-size 256
# Algorithm
--advantage-estimator grpo # or: gspo, ppo, reinforce_plus_plus
--use-kl-loss
--kl-loss-coef 0.001
```
### Key Constraints
```
rollout_batch_size × n_samples_per_prompt = global_batch_size × num_steps_per_rollout
```
Example: 32 × 8 = 256 × 1
---
## Data Buffer System
slime's data buffer enables flexible data management:
### Basic Data Source
```python
class RolloutDataSource:
def get_samples(self, num_samples):
"""Fetch prompts from dataset."""
return self.dataset.sample(num_samples)
def add_samples(self, samples):
"""Called after generation (no-op by default)."""
pass
```
### Buffered Data Source (Off-Policy)
```python
class RolloutDataSourceWithBuffer(RolloutDataSource):
def __init__(self):
self.buffer = []
def add_samples(self, samples):
"""Store generated samples for reuse."""
self.buffer.extend(samples)
def buffer_filter(self, args, buffer, num_samples):
"""Custom selection logic (prioritized, stratified, etc.)."""
return select_best(buffer, num_samples)
```
---
## Common Issues and Solutions
### Issue: SGLang Engine Crash
**Symptoms**: Inference engine dies mid-training
**Solutions**:
```bash
# Enable fault tolerance
--use-fault-tolerance
# Increase memory allocation
--sglang-mem-fraction-static 0.85
# Reduce batch size
--rollout-batch-size 16
```
### Issue: Weight Sync Timeout
**Symptoms**: Training hangs after rollout
**Solutions**:
```bash
# Increase sync interval
--update-weights-interval 5
# Use colocated mode (no network transfer)
--colocate
```
### Issue: OOM During Training
**Symptoms**: CUDA OOM in backward pass
**Solutions**:
```bash
# Enable gradient checkpointing
--recompute-activations
# Reduce micro-batch size
--micro-batch-size 1
# Enable sequence parallelism
--sequence-parallel
```
### Issue: Slow Data Loading
**Symptoms**: GPU idle during data fetch
**Solutions**:
```bash
# Increase data workers
--num-data-workers 4
# Use streaming dataset
--streaming-data
```
---
## Supported Models
| Model Family | Configurations |
|--------------|----------------|
| GLM | GLM-4.5, GLM-4.6, GLM-4.7, GLM-Z1-9B |
| Qwen | Qwen3 (4B, 8B, 30B-A3B), Qwen3-MoE, Qwen2.5 |
| DeepSeek | V3, V3.1, R1 |
| Llama | Llama 3 (8B, 70B) |
| Others | Kimi K2, Moonlight-16B |
Each model has pre-configured scripts in `scripts/models/`.
---
## Advanced Topics
### Co-location Mode
Share GPUs between training and inference to reduce memory:
```bash
python train.py \
--colocate \
--actor-num-gpus-per-node 8 \
--sglang-mem-fraction-static 0.4 \
${MODEL_ARGS[@]}
```
### Custom Reward Model
```python
# custom_rm.py
class CustomRewardModel:
def __init__(self, model_path):
self.model = load_model(model_path)
def compute_reward(self, prompts, responses):
inputs = self.tokenize(prompts, responses)
scores = self.model(inputs)
return scores.tolist()
```
```bash
--custom-rm-path custom_rm.py
```
### Evaluation Multi-Task
```bash
--eval-prompt-data aime /path/to/aime.jsonl \
--eval-prompt-data gsm8k /path/to/gsm8k.jsonl \
--n-samples-per-eval-prompt 16
```
---
## Resources
- **Documentation**: https://thudm.github.io/slime/
- **GitHub**: https://github.com/THUDM/slime
- **Blog**: https://lmsys.org/blog/2025-07-09-slime/
- **Examples**: See `examples/` directory for 14+ worked examples

539
website/docs/user-guide/skills/optional/mlops/mlops-stable-diffusion.md Normal file
View file

@ -0,0 +1,539 @@
---
title: "Stable Diffusion Image Generation"
sidebar_label: "Stable Diffusion Image Generation"
description: "State-of-the-art text-to-image generation with Stable Diffusion models via HuggingFace Diffusers"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Stable Diffusion Image Generation
State-of-the-art text-to-image generation with Stable Diffusion models via HuggingFace Diffusers. Use when generating images from text prompts, performing image-to-image translation, inpainting, or building custom diffusion pipelines.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/stable-diffusion` |
| Path | `optional-skills/mlops/stable-diffusion` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `diffusers>=0.30.0`, `transformers>=4.41.0`, `accelerate>=0.31.0`, `torch>=2.0.0` |
| Tags | `Image Generation`, `Stable Diffusion`, `Diffusers`, `Text-to-Image`, `Multimodal`, `Computer Vision` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Stable Diffusion Image Generation
Comprehensive guide to generating images with Stable Diffusion using the HuggingFace Diffusers library.
## When to use Stable Diffusion
**Use Stable Diffusion when:**
- Generating images from text descriptions
- Performing image-to-image translation (style transfer, enhancement)
- Inpainting (filling in masked regions)
- Outpainting (extending images beyond boundaries)
- Creating variations of existing images
- Building custom image generation workflows
**Key features:**
- **Text-to-Image**: Generate images from natural language prompts
- **Image-to-Image**: Transform existing images with text guidance
- **Inpainting**: Fill masked regions with context-aware content
- **ControlNet**: Add spatial conditioning (edges, poses, depth)
- **LoRA Support**: Efficient fine-tuning and style adaptation
- **Multiple Models**: SD 1.5, SDXL, SD 3.0, Flux support
**Use alternatives instead:**
- **DALL-E 3**: For API-based generation without GPU
- **Midjourney**: For artistic, stylized outputs
- **Imagen**: For Google Cloud integration
- **Leonardo.ai**: For web-based creative workflows
## Quick start
### Installation
```bash
pip install diffusers transformers accelerate torch
pip install xformers # Optional: memory-efficient attention
```
### Basic text-to-image
```python
from diffusers import DiffusionPipeline
import torch
# Load pipeline (auto-detects model type)
pipe = DiffusionPipeline.from_pretrained(
"stable-diffusion-v1-5/stable-diffusion-v1-5",
torch_dtype=torch.float16
)
pipe.to("cuda")
# Generate image
image = pipe(
"A serene mountain landscape at sunset, highly detailed",
num_inference_steps=50,
guidance_scale=7.5
).images[0]
image.save("output.png")
```
### Using SDXL (higher quality)
```python
from diffusers import AutoPipelineForText2Image
import torch
pipe = AutoPipelineForText2Image.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
torch_dtype=torch.float16,
variant="fp16"
)
pipe.to("cuda")
# Enable memory optimization
pipe.enable_model_cpu_offload()
image = pipe(
prompt="A futuristic city with flying cars, cinematic lighting",
height=1024,
width=1024,
num_inference_steps=30
).images[0]
```
## Architecture overview
### Three-pillar design
Diffusers is built around three core components:
```
Pipeline (orchestration)
├── Model (neural networks)
│ ├── UNet / Transformer (noise prediction)
│ ├── VAE (latent encoding/decoding)
│ └── Text Encoder (CLIP/T5)
└── Scheduler (denoising algorithm)
```
### Pipeline inference flow
```
Text Prompt → Text Encoder → Text Embeddings
↓
Random Noise → [Denoising Loop] ← Scheduler
↓
Predicted Noise
↓
VAE Decoder → Final Image
```
## Core concepts
### Pipelines
Pipelines orchestrate complete workflows:
| Pipeline | Purpose |
|----------|---------|
| `StableDiffusionPipeline` | Text-to-image (SD 1.x/2.x) |
| `StableDiffusionXLPipeline` | Text-to-image (SDXL) |
| `StableDiffusion3Pipeline` | Text-to-image (SD 3.0) |
| `FluxPipeline` | Text-to-image (Flux models) |
| `StableDiffusionImg2ImgPipeline` | Image-to-image |
| `StableDiffusionInpaintPipeline` | Inpainting |
### Schedulers
Schedulers control the denoising process:
| Scheduler | Steps | Quality | Use Case |
|-----------|-------|---------|----------|
| `EulerDiscreteScheduler` | 20-50 | Good | Default choice |
| `EulerAncestralDiscreteScheduler` | 20-50 | Good | More variation |
| `DPMSolverMultistepScheduler` | 15-25 | Excellent | Fast, high quality |
| `DDIMScheduler` | 50-100 | Good | Deterministic |
| `LCMScheduler` | 4-8 | Good | Very fast |
| `UniPCMultistepScheduler` | 15-25 | Excellent | Fast convergence |
### Swapping schedulers
```python
from diffusers import DPMSolverMultistepScheduler
# Swap for faster generation
pipe.scheduler = DPMSolverMultistepScheduler.from_config(
pipe.scheduler.config
)
# Now generate with fewer steps
image = pipe(prompt, num_inference_steps=20).images[0]
```
## Generation parameters
### Key parameters
| Parameter | Default | Description |
|-----------|---------|-------------|
| `prompt` | Required | Text description of desired image |
| `negative_prompt` | None | What to avoid in the image |
| `num_inference_steps` | 50 | Denoising steps (more = better quality) |
| `guidance_scale` | 7.5 | Prompt adherence (7-12 typical) |
| `height`, `width` | 512/1024 | Output dimensions (multiples of 8) |
| `generator` | None | Torch generator for reproducibility |
| `num_images_per_prompt` | 1 | Batch size |
### Reproducible generation
```python
import torch
generator = torch.Generator(device="cuda").manual_seed(42)
image = pipe(
prompt="A cat wearing a top hat",
generator=generator,
num_inference_steps=50
).images[0]
```
### Negative prompts
```python
image = pipe(
prompt="Professional photo of a dog in a garden",
negative_prompt="blurry, low quality, distorted, ugly, bad anatomy",
guidance_scale=7.5
).images[0]
```
## Image-to-image
Transform existing images with text guidance:
```python
from diffusers import AutoPipelineForImage2Image
from PIL import Image
pipe = AutoPipelineForImage2Image.from_pretrained(
"stable-diffusion-v1-5/stable-diffusion-v1-5",
torch_dtype=torch.float16
).to("cuda")
init_image = Image.open("input.jpg").resize((512, 512))
image = pipe(
prompt="A watercolor painting of the scene",
image=init_image,
strength=0.75, # How much to transform (0-1)
num_inference_steps=50
).images[0]
```
## Inpainting
Fill masked regions:
```python
from diffusers import AutoPipelineForInpainting
from PIL import Image
pipe = AutoPipelineForInpainting.from_pretrained(
"runwayml/stable-diffusion-inpainting",
torch_dtype=torch.float16
).to("cuda")
image = Image.open("photo.jpg")
mask = Image.open("mask.png") # White = inpaint region
result = pipe(
prompt="A red car parked on the street",
image=image,
mask_image=mask,
num_inference_steps=50
).images[0]
```
## ControlNet
Add spatial conditioning for precise control:
```python
from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
import torch
# Load ControlNet for edge conditioning
controlnet = ControlNetModel.from_pretrained(
"lllyasviel/control_v11p_sd15_canny",
torch_dtype=torch.float16
)
pipe = StableDiffusionControlNetPipeline.from_pretrained(
"stable-diffusion-v1-5/stable-diffusion-v1-5",
controlnet=controlnet,
torch_dtype=torch.float16
).to("cuda")
# Use Canny edge image as control
control_image = get_canny_image(input_image)
image = pipe(
prompt="A beautiful house in the style of Van Gogh",
image=control_image,
num_inference_steps=30
).images[0]
```
### Available ControlNets
| ControlNet | Input Type | Use Case |
|------------|------------|----------|
| `canny` | Edge maps | Preserve structure |
| `openpose` | Pose skeletons | Human poses |
| `depth` | Depth maps | 3D-aware generation |
| `normal` | Normal maps | Surface details |
| `mlsd` | Line segments | Architectural lines |
| `scribble` | Rough sketches | Sketch-to-image |
## LoRA adapters
Load fine-tuned style adapters:
```python
from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained(
"stable-diffusion-v1-5/stable-diffusion-v1-5",
torch_dtype=torch.float16
).to("cuda")
# Load LoRA weights
pipe.load_lora_weights("path/to/lora", weight_name="style.safetensors")
# Generate with LoRA style
image = pipe("A portrait in the trained style").images[0]
# Adjust LoRA strength
pipe.fuse_lora(lora_scale=0.8)
# Unload LoRA
pipe.unload_lora_weights()
```
### Multiple LoRAs
```python
# Load multiple LoRAs
pipe.load_lora_weights("lora1", adapter_name="style")
pipe.load_lora_weights("lora2", adapter_name="character")
# Set weights for each
pipe.set_adapters(["style", "character"], adapter_weights=[0.7, 0.5])
image = pipe("A portrait").images[0]
```
## Memory optimization
### Enable CPU offloading
```python
# Model CPU offload - moves models to CPU when not in use
pipe.enable_model_cpu_offload()
# Sequential CPU offload - more aggressive, slower
pipe.enable_sequential_cpu_offload()
```
### Attention slicing
```python
# Reduce memory by computing attention in chunks
pipe.enable_attention_slicing()
# Or specific chunk size
pipe.enable_attention_slicing("max")
```
### xFormers memory-efficient attention
```python
# Requires xformers package
pipe.enable_xformers_memory_efficient_attention()
```
### VAE slicing for large images
```python
# Decode latents in tiles for large images
pipe.enable_vae_slicing()
pipe.enable_vae_tiling()
```
## Model variants
### Loading different precisions
```python
# FP16 (recommended for GPU)
pipe = DiffusionPipeline.from_pretrained(
"model-id",
torch_dtype=torch.float16,
variant="fp16"
)
# BF16 (better precision, requires Ampere+ GPU)
pipe = DiffusionPipeline.from_pretrained(
"model-id",
torch_dtype=torch.bfloat16
)
```
### Loading specific components
```python
from diffusers import UNet2DConditionModel, AutoencoderKL
# Load custom VAE
vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse")
# Use with pipeline
pipe = DiffusionPipeline.from_pretrained(
"stable-diffusion-v1-5/stable-diffusion-v1-5",
vae=vae,
torch_dtype=torch.float16
)
```
## Batch generation
Generate multiple images efficiently:
```python
# Multiple prompts
prompts = [
"A cat playing piano",
"A dog reading a book",
"A bird painting a picture"
]
images = pipe(prompts, num_inference_steps=30).images
# Multiple images per prompt
images = pipe(
"A beautiful sunset",
num_images_per_prompt=4,
num_inference_steps=30
).images
```
## Common workflows
### Workflow 1: High-quality generation
```python
from diffusers import StableDiffusionXLPipeline, DPMSolverMultistepScheduler
import torch
# 1. Load SDXL with optimizations
pipe = StableDiffusionXLPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
torch_dtype=torch.float16,
variant="fp16"
)
pipe.to("cuda")
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
pipe.enable_model_cpu_offload()
# 2. Generate with quality settings
image = pipe(
prompt="A majestic lion in the savanna, golden hour lighting, 8k, detailed fur",
negative_prompt="blurry, low quality, cartoon, anime, sketch",
num_inference_steps=30,
guidance_scale=7.5,
height=1024,
width=1024
).images[0]
```
### Workflow 2: Fast prototyping
```python
from diffusers import AutoPipelineForText2Image, LCMScheduler
import torch
# Use LCM for 4-8 step generation
pipe = AutoPipelineForText2Image.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
torch_dtype=torch.float16
).to("cuda")
# Load LCM LoRA for fast generation
pipe.load_lora_weights("latent-consistency/lcm-lora-sdxl")
pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
pipe.fuse_lora()
# Generate in ~1 second
image = pipe(
"A beautiful landscape",
num_inference_steps=4,
guidance_scale=1.0
).images[0]
```
## Common issues
**CUDA out of memory:**
```python
# Enable memory optimizations
pipe.enable_model_cpu_offload()
pipe.enable_attention_slicing()
pipe.enable_vae_slicing()
# Or use lower precision
pipe = DiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
```
**Black/noise images:**
```python
# Check VAE configuration
# Use safety checker bypass if needed
pipe.safety_checker = None
# Ensure proper dtype consistency
pipe = pipe.to(dtype=torch.float16)
```
**Slow generation:**
```python
# Use faster scheduler
from diffusers import DPMSolverMultistepScheduler
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
# Reduce steps
image = pipe(prompt, num_inference_steps=20).images[0]
```
## References
- **[Advanced Usage](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/stable-diffusion/references/advanced-usage.md)** - Custom pipelines, fine-tuning, deployment
- **[Troubleshooting](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/stable-diffusion/references/troubleshooting.md)** - Common issues and solutions
## Resources
- **Documentation**: https://huggingface.co/docs/diffusers
- **Repository**: https://github.com/huggingface/diffusers
- **Model Hub**: https://huggingface.co/models?library=diffusers
- **Discord**: https://discord.gg/diffusers

205
website/docs/user-guide/skills/optional/mlops/mlops-tensorrt-llm.md Normal file
View file

@ -0,0 +1,205 @@
---
title: "Tensorrt Llm — Optimizes LLM inference with NVIDIA TensorRT for maximum throughput and lowest latency"
sidebar_label: "Tensorrt Llm"
description: "Optimizes LLM inference with NVIDIA TensorRT for maximum throughput and lowest latency"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Tensorrt Llm
Optimizes LLM inference with NVIDIA TensorRT for maximum throughput and lowest latency. Use for production deployment on NVIDIA GPUs (A100/H100), when you need 10-100x faster inference than PyTorch, or for serving models with quantization (FP8/INT4), in-flight batching, and multi-GPU scaling.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/tensorrt-llm` |
| Path | `optional-skills/mlops/tensorrt-llm` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `tensorrt-llm`, `torch` |
| Tags | `Inference Serving`, `TensorRT-LLM`, `NVIDIA`, `Inference Optimization`, `High Throughput`, `Low Latency`, `Production`, `FP8`, `INT4`, `In-Flight Batching`, `Multi-GPU` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# TensorRT-LLM
NVIDIA's open-source library for optimizing LLM inference with state-of-the-art performance on NVIDIA GPUs.
## When to use TensorRT-LLM
**Use TensorRT-LLM when:**
- Deploying on NVIDIA GPUs (A100, H100, GB200)
- Need maximum throughput (24,000+ tokens/sec on Llama 3)
- Require low latency for real-time applications
- Working with quantized models (FP8, INT4, FP4)
- Scaling across multiple GPUs or nodes
**Use vLLM instead when:**
- Need simpler setup and Python-first API
- Want PagedAttention without TensorRT compilation
- Working with AMD GPUs or non-NVIDIA hardware
**Use llama.cpp instead when:**
- Deploying on CPU or Apple Silicon
- Need edge deployment without NVIDIA GPUs
- Want simpler GGUF quantization format
## Quick start
### Installation
```bash
# Docker (recommended)
docker pull nvidia/tensorrt_llm:latest
# pip install
pip install tensorrt_llm==1.2.0rc3
# Requires CUDA 13.0.0, TensorRT 10.13.2, Python 3.10-3.12
```
### Basic inference
```python
from tensorrt_llm import LLM, SamplingParams
# Initialize model
llm = LLM(model="meta-llama/Meta-Llama-3-8B")
# Configure sampling
sampling_params = SamplingParams(
max_tokens=100,
temperature=0.7,
top_p=0.9
)
# Generate
prompts = ["Explain quantum computing"]
outputs = llm.generate(prompts, sampling_params)
for output in outputs:
print(output.text)
```
### Serving with trtllm-serve
```bash
# Start server (automatic model download and compilation)
trtllm-serve meta-llama/Meta-Llama-3-8B \
--tp_size 4 \ # Tensor parallelism (4 GPUs)
--max_batch_size 256 \
--max_num_tokens 4096
# Client request
curl -X POST http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "meta-llama/Meta-Llama-3-8B",
"messages": [{"role": "user", "content": "Hello!"}],
"temperature": 0.7,
"max_tokens": 100
}'
```
## Key features
### Performance optimizations
- **In-flight batching**: Dynamic batching during generation
- **Paged KV cache**: Efficient memory management
- **Flash Attention**: Optimized attention kernels
- **Quantization**: FP8, INT4, FP4 for 2-4× faster inference
- **CUDA graphs**: Reduced kernel launch overhead
### Parallelism
- **Tensor parallelism (TP)**: Split model across GPUs
- **Pipeline parallelism (PP)**: Layer-wise distribution
- **Expert parallelism**: For Mixture-of-Experts models
- **Multi-node**: Scale beyond single machine
### Advanced features
- **Speculative decoding**: Faster generation with draft models
- **LoRA serving**: Efficient multi-adapter deployment
- **Disaggregated serving**: Separate prefill and generation
## Common patterns
### Quantized model (FP8)
```python
from tensorrt_llm import LLM
# Load FP8 quantized model (2× faster, 50% memory)
llm = LLM(
model="meta-llama/Meta-Llama-3-70B",
dtype="fp8",
max_num_tokens=8192
)
# Inference same as before
outputs = llm.generate(["Summarize this article..."])
```
### Multi-GPU deployment
```python
# Tensor parallelism across 8 GPUs
llm = LLM(
model="meta-llama/Meta-Llama-3-405B",
tensor_parallel_size=8,
dtype="fp8"
)
```
### Batch inference
```python
# Process 100 prompts efficiently
prompts = [f"Question {i}: ..." for i in range(100)]
outputs = llm.generate(
prompts,
sampling_params=SamplingParams(max_tokens=200)
)
# Automatic in-flight batching for maximum throughput
```
## Performance benchmarks
**Meta Llama 3-8B** (H100 GPU):
- Throughput: 24,000 tokens/sec
- Latency: ~10ms per token
- vs PyTorch: **100× faster**
**Llama 3-70B** (8× A100 80GB):
- FP8 quantization: 2× faster than FP16
- Memory: 50% reduction with FP8
## Supported models
- **LLaMA family**: Llama 2, Llama 3, CodeLlama
- **GPT family**: GPT-2, GPT-J, GPT-NeoX
- **Qwen**: Qwen, Qwen2, QwQ
- **DeepSeek**: DeepSeek-V2, DeepSeek-V3
- **Mixtral**: Mixtral-8x7B, Mixtral-8x22B
- **Vision**: LLaVA, Phi-3-vision
- **100+ models** on HuggingFace
## References
- **[Optimization Guide](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/tensorrt-llm/references/optimization.md)** - Quantization, batching, KV cache tuning
- **[Multi-GPU Setup](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/tensorrt-llm/references/multi-gpu.md)** - Tensor/pipeline parallelism, multi-node
- **[Serving Guide](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/tensorrt-llm/references/serving.md)** - Production deployment, monitoring, autoscaling
## Resources
- **Docs**: https://nvidia.github.io/TensorRT-LLM/
- **GitHub**: https://github.com/NVIDIA/TensorRT-LLM
- **Models**: https://huggingface.co/models?library=tensorrt_llm

377
website/docs/user-guide/skills/optional/mlops/mlops-torchtitan.md Normal file
View file

@ -0,0 +1,377 @@
---
title: "Distributed Llm Pretraining Torchtitan"
sidebar_label: "Distributed Llm Pretraining Torchtitan"
description: "Provides PyTorch-native distributed LLM pretraining using torchtitan with 4D parallelism (FSDP2, TP, PP, CP)"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Distributed Llm Pretraining Torchtitan
Provides PyTorch-native distributed LLM pretraining using torchtitan with 4D parallelism (FSDP2, TP, PP, CP). Use when pretraining Llama 3.1, DeepSeek V3, or custom models at scale from 8 to 512+ GPUs with Float8, torch.compile, and distributed checkpointing.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/torchtitan` |
| Path | `optional-skills/mlops/torchtitan` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `torch>=2.6.0`, `torchtitan>=0.2.0`, `torchao>=0.5.0` |
| Tags | `Model Architecture`, `Distributed Training`, `TorchTitan`, `FSDP2`, `Tensor Parallel`, `Pipeline Parallel`, `Context Parallel`, `Float8`, `Llama`, `Pretraining` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# TorchTitan - PyTorch Native Distributed LLM Pretraining
## Quick start
TorchTitan is PyTorch's official platform for large-scale LLM pretraining with composable 4D parallelism (FSDP2, TP, PP, CP), achieving 65%+ speedups over baselines on H100 GPUs.
**Installation**:
```bash
# From PyPI (stable)
pip install torchtitan
# From source (latest features, requires PyTorch nightly)
git clone https://github.com/pytorch/torchtitan
cd torchtitan
pip install -r requirements.txt
```
**Download tokenizer**:
```bash
# Get HF token from https://huggingface.co/settings/tokens
python scripts/download_hf_assets.py --repo_id meta-llama/Llama-3.1-8B --assets tokenizer --hf_token=...
```
**Start training on 8 GPUs**:
```bash
CONFIG_FILE="./torchtitan/models/llama3/train_configs/llama3_8b.toml" ./run_train.sh
```
## Common workflows
### Workflow 1: Pretrain Llama 3.1 8B on single node
Copy this checklist:
```
Single Node Pretraining:
- [ ] Step 1: Download tokenizer
- [ ] Step 2: Configure training
- [ ] Step 3: Launch training
- [ ] Step 4: Monitor and checkpoint
```
**Step 1: Download tokenizer**
```bash
python scripts/download_hf_assets.py \
--repo_id meta-llama/Llama-3.1-8B \
--assets tokenizer \
--hf_token=YOUR_HF_TOKEN
```
**Step 2: Configure training**
Edit or create a TOML config file:
```toml
# llama3_8b_custom.toml
[job]
dump_folder = "./outputs"
description = "Llama 3.1 8B training"
[model]
name = "llama3"
flavor = "8B"
hf_assets_path = "./assets/hf/Llama-3.1-8B"
[optimizer]
name = "AdamW"
lr = 3e-4
[lr_scheduler]
warmup_steps = 200
[training]
local_batch_size = 2
seq_len = 8192
max_norm = 1.0
steps = 1000
dataset = "c4"
[parallelism]
data_parallel_shard_degree = -1 # Use all GPUs for FSDP
[activation_checkpoint]
mode = "selective"
selective_ac_option = "op"
[checkpoint]
enable = true
folder = "checkpoint"
interval = 500
```
**Step 3: Launch training**
```bash
# 8 GPUs on single node
CONFIG_FILE="./llama3_8b_custom.toml" ./run_train.sh
# Or explicitly with torchrun
torchrun --nproc_per_node=8 \
-m torchtitan.train \
--job.config_file ./llama3_8b_custom.toml
```
**Step 4: Monitor and checkpoint**
TensorBoard logs are saved to `./outputs/tb/`:
```bash
tensorboard --logdir ./outputs/tb
```
### Workflow 2: Multi-node training with SLURM
```
Multi-Node Training:
- [ ] Step 1: Configure parallelism for scale
- [ ] Step 2: Set up SLURM script
- [ ] Step 3: Submit job
- [ ] Step 4: Resume from checkpoint
```
**Step 1: Configure parallelism for scale**
For 70B model on 256 GPUs (32 nodes):
```toml
[parallelism]
data_parallel_shard_degree = 32 # FSDP across 32 ranks
tensor_parallel_degree = 8 # TP within node
pipeline_parallel_degree = 1 # No PP for 70B
context_parallel_degree = 1 # Increase for long sequences
```
**Step 2: Set up SLURM script**
```bash
#!/bin/bash
#SBATCH --job-name=llama70b
#SBATCH --nodes=32
#SBATCH --ntasks-per-node=8
#SBATCH --gpus-per-node=8
srun torchrun \
--nnodes=32 \
--nproc_per_node=8 \
--rdzv_backend=c10d \
--rdzv_endpoint=$MASTER_ADDR:$MASTER_PORT \
-m torchtitan.train \
--job.config_file ./llama3_70b.toml
```
**Step 3: Submit job**
```bash
sbatch multinode_trainer.slurm
```
**Step 4: Resume from checkpoint**
Training auto-resumes if checkpoint exists in configured folder.
### Workflow 3: Enable Float8 training for H100s
Float8 provides 30-50% speedup on H100 GPUs.
```
Float8 Training:
- [ ] Step 1: Install torchao
- [ ] Step 2: Configure Float8
- [ ] Step 3: Launch with compile
```
**Step 1: Install torchao**
```bash
USE_CPP=0 pip install git+https://github.com/pytorch/ao.git
```
**Step 2: Configure Float8**
Add to your TOML config:
```toml
[model]
converters = ["quantize.linear.float8"]
[quantize.linear.float8]
enable_fsdp_float8_all_gather = true
precompute_float8_dynamic_scale_for_fsdp = true
filter_fqns = ["output"] # Exclude output layer
[compile]
enable = true
components = ["model", "loss"]
```
**Step 3: Launch with compile**
```bash
CONFIG_FILE="./llama3_8b.toml" ./run_train.sh \
--model.converters="quantize.linear.float8" \
--quantize.linear.float8.enable_fsdp_float8_all_gather \
--compile.enable
```
### Workflow 4: 4D parallelism for 405B models
```
4D Parallelism (FSDP + TP + PP + CP):
- [ ] Step 1: Create seed checkpoint
- [ ] Step 2: Configure 4D parallelism
- [ ] Step 3: Launch on 512 GPUs
```
**Step 1: Create seed checkpoint**
Required for consistent initialization across PP stages:
```bash
NGPU=1 CONFIG_FILE=./llama3_405b.toml ./run_train.sh \
--checkpoint.enable \
--checkpoint.create_seed_checkpoint \
--parallelism.data_parallel_shard_degree 1 \
--parallelism.tensor_parallel_degree 1 \
--parallelism.pipeline_parallel_degree 1
```
**Step 2: Configure 4D parallelism**
```toml
[parallelism]
data_parallel_shard_degree = 8 # FSDP
tensor_parallel_degree = 8 # TP within node
pipeline_parallel_degree = 8 # PP across nodes
context_parallel_degree = 1 # CP for long sequences
[training]
local_batch_size = 32
seq_len = 8192
```
**Step 3: Launch on 512 GPUs**
```bash
# 64 nodes x 8 GPUs = 512 GPUs
srun torchrun --nnodes=64 --nproc_per_node=8 \
-m torchtitan.train \
--job.config_file ./llama3_405b.toml
```
## When to use vs alternatives
**Use TorchTitan when:**
- Pretraining LLMs from scratch (8B to 405B+)
- Need PyTorch-native solution without third-party dependencies
- Require composable 4D parallelism (FSDP2, TP, PP, CP)
- Training on H100s with Float8 support
- Want interoperable checkpoints with torchtune/HuggingFace
**Use alternatives instead:**
- **Megatron-LM**: Maximum performance for NVIDIA-only deployments
- **DeepSpeed**: Broader ZeRO optimization ecosystem, inference support
- **Axolotl/TRL**: Fine-tuning rather than pretraining
- **LitGPT**: Educational, smaller-scale training
## Common issues
**Issue: Out of memory on large models**
Enable activation checkpointing and reduce batch size:
```toml
[activation_checkpoint]
mode = "full" # Instead of "selective"
[training]
local_batch_size = 1
```
Or use gradient accumulation:
```toml
[training]
local_batch_size = 1
global_batch_size = 32 # Accumulates gradients
```
**Issue: TP causes high memory with async collectives**
Set environment variable:
```bash
export TORCH_NCCL_AVOID_RECORD_STREAMS=1
```
**Issue: Float8 training not faster**
Float8 only benefits large GEMMs. Filter small layers:
```toml
[quantize.linear.float8]
filter_fqns = ["attention.wk", "attention.wv", "output", "auto_filter_small_kn"]
```
**Issue: Checkpoint loading fails after parallelism change**
Use DCP's resharding capability:
```bash
# Convert sharded checkpoint to single file
python -m torch.distributed.checkpoint.format_utils \
dcp_to_torch checkpoint/step-1000 checkpoint.pt
```
**Issue: Pipeline parallelism initialization**
Create seed checkpoint first (see Workflow 4, Step 1).
## Supported models
| Model | Sizes | Status |
|-------|-------|--------|
| Llama 3.1 | 8B, 70B, 405B | Production |
| Llama 4 | Various | Experimental |
| DeepSeek V3 | 16B, 236B, 671B (MoE) | Experimental |
| GPT-OSS | 20B, 120B (MoE) | Experimental |
| Qwen 3 | Various | Experimental |
| Flux | Diffusion | Experimental |
## Performance benchmarks (H100)
| Model | GPUs | Parallelism | TPS/GPU | Techniques |
|-------|------|-------------|---------|------------|
| Llama 8B | 8 | FSDP | 5,762 | Baseline |
| Llama 8B | 8 | FSDP+compile+FP8 | 8,532 | +48% |
| Llama 70B | 256 | FSDP+TP+AsyncTP | 876 | 2D parallel |
| Llama 405B | 512 | FSDP+TP+PP | 128 | 3D parallel |
## Advanced topics
**FSDP2 configuration**: See [references/fsdp.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/torchtitan/references/fsdp.md) for detailed FSDP2 vs FSDP1 comparison and ZeRO equivalents.
**Float8 training**: See [references/float8.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/torchtitan/references/float8.md) for tensorwise vs rowwise scaling recipes.
**Checkpointing**: See [references/checkpoint.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/torchtitan/references/checkpoint.md) for HuggingFace conversion and async checkpointing.
**Adding custom models**: See [references/custom-models.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops/torchtitan/references/custom-models.md) for TrainSpec protocol.
## Resources
- GitHub: https://github.com/pytorch/torchtitan
- Paper: https://arxiv.org/abs/2410.06511
- ICLR 2025: https://iclr.cc/virtual/2025/poster/29620
- PyTorch Forum: https://discuss.pytorch.org/c/distributed/torchtitan/44

335
website/docs/user-guide/skills/optional/mlops/mlops-whisper.md Normal file
View file

@ -0,0 +1,335 @@
---
title: "Whisper — OpenAI's general-purpose speech recognition model"
sidebar_label: "Whisper"
description: "OpenAI's general-purpose speech recognition model"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Whisper
OpenAI's general-purpose speech recognition model. Supports 99 languages, transcription, translation to English, and language identification. Six model sizes from tiny (39M params) to large (1550M params). Use for speech-to-text, podcast transcription, or multilingual audio processing. Best for robust, multilingual ASR.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/whisper` |
| Path | `optional-skills/mlops/whisper` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `openai-whisper`, `transformers`, `torch` |
| Tags | `Whisper`, `Speech Recognition`, `ASR`, `Multimodal`, `Multilingual`, `OpenAI`, `Speech-To-Text`, `Transcription`, `Translation`, `Audio Processing` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Whisper - Robust Speech Recognition
OpenAI's multilingual speech recognition model.
## When to use Whisper
**Use when:**
- Speech-to-text transcription (99 languages)
- Podcast/video transcription
- Meeting notes automation
- Translation to English
- Noisy audio transcription
- Multilingual audio processing
**Metrics**:
- **72,900+ GitHub stars**
- 99 languages supported
- Trained on 680,000 hours of audio
- MIT License
**Use alternatives instead**:
- **AssemblyAI**: Managed API, speaker diarization
- **Deepgram**: Real-time streaming ASR
- **Google Speech-to-Text**: Cloud-based
## Quick start
### Installation
```bash
# Requires Python 3.8-3.11
pip install -U openai-whisper
# Requires ffmpeg
# macOS: brew install ffmpeg
# Ubuntu: sudo apt install ffmpeg
# Windows: choco install ffmpeg
```
### Basic transcription
```python
import whisper
# Load model
model = whisper.load_model("base")
# Transcribe
result = model.transcribe("audio.mp3")
# Print text
print(result["text"])
# Access segments
for segment in result["segments"]:
print(f"[{segment['start']:.2f}s - {segment['end']:.2f}s] {segment['text']}")
```
## Model sizes
```python
# Available models
models = ["tiny", "base", "small", "medium", "large", "turbo"]
# Load specific model
model = whisper.load_model("turbo") # Fastest, good quality
```
| Model | Parameters | English-only | Multilingual | Speed | VRAM |
|-------|------------|--------------|--------------|-------|------|
| tiny | 39M | ✓ | ✓ | ~32x | ~1 GB |
| base | 74M | ✓ | ✓ | ~16x | ~1 GB |
| small | 244M | ✓ | ✓ | ~6x | ~2 GB |
| medium | 769M | ✓ | ✓ | ~2x | ~5 GB |
| large | 1550M | ✗ | ✓ | 1x | ~10 GB |
| turbo | 809M | ✗ | ✓ | ~8x | ~6 GB |
**Recommendation**: Use `turbo` for best speed/quality, `base` for prototyping
## Transcription options
### Language specification
```python
# Auto-detect language
result = model.transcribe("audio.mp3")
# Specify language (faster)
result = model.transcribe("audio.mp3", language="en")
# Supported: en, es, fr, de, it, pt, ru, ja, ko, zh, and 89 more
```
### Task selection
```python
# Transcription (default)
result = model.transcribe("audio.mp3", task="transcribe")
# Translation to English
result = model.transcribe("spanish.mp3", task="translate")
# Input: Spanish audio → Output: English text
```
### Initial prompt
```python
# Improve accuracy with context
result = model.transcribe(
"audio.mp3",
initial_prompt="This is a technical podcast about machine learning and AI."
)
# Helps with:
# - Technical terms
# - Proper nouns
# - Domain-specific vocabulary
```
### Timestamps
```python
# Word-level timestamps
result = model.transcribe("audio.mp3", word_timestamps=True)
for segment in result["segments"]:
for word in segment["words"]:
print(f"{word['word']} ({word['start']:.2f}s - {word['end']:.2f}s)")
```
### Temperature fallback
```python
# Retry with different temperatures if confidence low
result = model.transcribe(
"audio.mp3",
temperature=(0.0, 0.2, 0.4, 0.6, 0.8, 1.0)
)
```
## Command line usage
```bash
# Basic transcription
whisper audio.mp3
# Specify model
whisper audio.mp3 --model turbo
# Output formats
whisper audio.mp3 --output_format txt # Plain text
whisper audio.mp3 --output_format srt # Subtitles
whisper audio.mp3 --output_format vtt # WebVTT
whisper audio.mp3 --output_format json # JSON with timestamps
# Language
whisper audio.mp3 --language Spanish
# Translation
whisper spanish.mp3 --task translate
```
## Batch processing
```python
import os
audio_files = ["file1.mp3", "file2.mp3", "file3.mp3"]
for audio_file in audio_files:
print(f"Transcribing {audio_file}...")
result = model.transcribe(audio_file)
# Save to file
output_file = audio_file.replace(".mp3", ".txt")
with open(output_file, "w") as f:
f.write(result["text"])
```
## Real-time transcription
```python
# For streaming audio, use faster-whisper
# pip install faster-whisper
from faster_whisper import WhisperModel
model = WhisperModel("base", device="cuda", compute_type="float16")
# Transcribe with streaming
segments, info = model.transcribe("audio.mp3", beam_size=5)
for segment in segments:
print(f"[{segment.start:.2f}s -> {segment.end:.2f}s] {segment.text}")
```
## GPU acceleration
```python
import whisper
# Automatically uses GPU if available
model = whisper.load_model("turbo")
# Force CPU
model = whisper.load_model("turbo", device="cpu")
# Force GPU
model = whisper.load_model("turbo", device="cuda")
# 10-20× faster on GPU
```
## Integration with other tools
### Subtitle generation
```bash
# Generate SRT subtitles
whisper video.mp4 --output_format srt --language English
# Output: video.srt
```
### With LangChain
```python
from langchain.document_loaders import WhisperTranscriptionLoader
loader = WhisperTranscriptionLoader(file_path="audio.mp3")
docs = loader.load()
# Use transcription in RAG
from langchain_chroma import Chroma
from langchain_openai import OpenAIEmbeddings
vectorstore = Chroma.from_documents(docs, OpenAIEmbeddings())
```
### Extract audio from video
```bash
# Use ffmpeg to extract audio
ffmpeg -i video.mp4 -vn -acodec pcm_s16le audio.wav
# Then transcribe
whisper audio.wav
```
## Best practices
1. **Use turbo model** - Best speed/quality for English
2. **Specify language** - Faster than auto-detect
3. **Add initial prompt** - Improves technical terms
4. **Use GPU** - 10-20× faster
5. **Batch process** - More efficient
6. **Convert to WAV** - Better compatibility
7. **Split long audio** - &lt;30 min chunks
8. **Check language support** - Quality varies by language
9. **Use faster-whisper** - 4× faster than openai-whisper
10. **Monitor VRAM** - Scale model size to hardware
## Performance
| Model | Real-time factor (CPU) | Real-time factor (GPU) |
|-------|------------------------|------------------------|
| tiny | ~0.32 | ~0.01 |
| base | ~0.16 | ~0.01 |
| turbo | ~0.08 | ~0.01 |
| large | ~1.0 | ~0.05 |
*Real-time factor: 0.1 = 10× faster than real-time*
## Language support
Top-supported languages:
- English (en)
- Spanish (es)
- French (fr)
- German (de)
- Italian (it)
- Portuguese (pt)
- Russian (ru)
- Japanese (ja)
- Korean (ko)
- Chinese (zh)
Full list: 99 languages total
## Limitations
1. **Hallucinations** - May repeat or invent text
2. **Long-form accuracy** - Degrades on >30 min audio
3. **Speaker identification** - No diarization
4. **Accents** - Quality varies
5. **Background noise** - Can affect accuracy
6. **Real-time latency** - Not suitable for live captioning
## Resources
- **GitHub**: https://github.com/openai/whisper ⭐ 72,900+
- **Paper**: https://arxiv.org/abs/2212.04356
- **Model Card**: https://github.com/openai/whisper/blob/main/model-card.md
- **Colab**: Available in repo
- **License**: MIT

113
website/docs/user-guide/skills/optional/productivity/productivity-canvas.md Normal file
View file

@ -0,0 +1,113 @@
---
title: "Canvas — Canvas LMS integration — fetch enrolled courses and assignments using API token authentication"
sidebar_label: "Canvas"
description: "Canvas LMS integration — fetch enrolled courses and assignments using API token authentication"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Canvas
Canvas LMS integration — fetch enrolled courses and assignments using API token authentication.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/productivity/canvas` |
| Path | `optional-skills/productivity/canvas` |
| Version | `1.0.0` |
| Author | community |
| License | MIT |
| Tags | `Canvas`, `LMS`, `Education`, `Courses`, `Assignments` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Canvas LMS — Course & Assignment Access
Read-only access to Canvas LMS for listing courses and assignments.
## Scripts
- `scripts/canvas_api.py` — Python CLI for Canvas API calls
## Setup
1. Log in to your Canvas instance in a browser
2. Go to **Account → Settings** (click your profile icon, then Settings)
3. Scroll to **Approved Integrations** and click **+ New Access Token**
4. Name the token (e.g., "Hermes Agent"), set an optional expiry, and click **Generate Token**
5. Copy the token and add to `~/.hermes/.env`:
```
CANVAS_API_TOKEN=your_token_here
CANVAS_BASE_URL=https://yourschool.instructure.com
```
The base URL is whatever appears in your browser when you're logged into Canvas (no trailing slash).
## Usage
```bash
CANVAS="python $HERMES_HOME/skills/productivity/canvas/scripts/canvas_api.py"
# List all active courses
$CANVAS list_courses --enrollment-state active
# List all courses (any state)
$CANVAS list_courses
# List assignments for a specific course
$CANVAS list_assignments 12345
# List assignments ordered by due date
$CANVAS list_assignments 12345 --order-by due_at
```
## Output Format
**list_courses** returns:
```json
[{"id": 12345, "name": "Intro to CS", "course_code": "CS101", "workflow_state": "available", "start_at": "...", "end_at": "..."}]
```
**list_assignments** returns:
```json
[{"id": 67890, "name": "Homework 1", "due_at": "2025-02-15T23:59:00Z", "points_possible": 100, "submission_types": ["online_upload"], "html_url": "...", "description": "...", "course_id": 12345}]
```
Note: Assignment descriptions are truncated to 500 characters. The `html_url` field links to the full assignment page in Canvas.
## API Reference (curl)
```bash
# List courses
curl -s -H "Authorization: Bearer $CANVAS_API_TOKEN" \
"$CANVAS_BASE_URL/api/v1/courses?enrollment_state=active&per_page=10"
# List assignments for a course
curl -s -H "Authorization: Bearer $CANVAS_API_TOKEN" \
"$CANVAS_BASE_URL/api/v1/courses/COURSE_ID/assignments?per_page=10&order_by=due_at"
```
Canvas uses `Link` headers for pagination. The Python script handles pagination automatically.
## Rules
- This skill is **read-only** — it only fetches data, never modifies courses or assignments
- On first use, verify auth by running `$CANVAS list_courses` — if it fails with 401, guide the user through setup
- Canvas rate-limits to ~700 requests per 10 minutes; check `X-Rate-Limit-Remaining` header if hitting limits
## Troubleshooting
| Problem | Fix |
|---------|-----|
| 401 Unauthorized | Token invalid or expired — regenerate in Canvas Settings |
| 403 Forbidden | Token lacks permission for this course |
| Empty course list | Try `--enrollment-state active` or omit the flag to see all states |
| Wrong institution | Verify `CANVAS_BASE_URL` matches the URL in your browser |
| Timeout errors | Check network connectivity to your Canvas instance |

336
website/docs/user-guide/skills/optional/productivity/productivity-memento-flashcards.md Normal file
View file

@ -0,0 +1,336 @@
---
title: "Memento Flashcards — Spaced-repetition flashcard system"
sidebar_label: "Memento Flashcards"
description: "Spaced-repetition flashcard system"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Memento Flashcards
Spaced-repetition flashcard system. Create cards from facts or text, chat with flashcards using free-text answers graded by the agent, generate quizzes from YouTube transcripts, review due cards with adaptive scheduling, and export/import decks as CSV.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/productivity/memento-flashcards` |
| Path | `optional-skills/productivity/memento-flashcards` |
| Version | `1.0.0` |
| Author | Memento AI |
| License | MIT |
| Platforms | macos, linux |
| Tags | `Education`, `Flashcards`, `Spaced Repetition`, `Learning`, `Quiz`, `YouTube` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Memento Flashcards — Spaced-Repetition Flashcard Skill
## Overview
Memento gives you a local, file-based flashcard system with spaced-repetition scheduling.
Users can chat with their flashcards by answering in free text and having the agent grade the response before scheduling the next review.
Use it whenever the user wants to:
- **Remember a fact** — turn any statement into a Q/A flashcard
- **Study with spaced repetition** — review due cards with adaptive intervals and agent-graded free-text answers
- **Quiz from a YouTube video** — fetch a transcript and generate a 5-question quiz
- **Manage decks** — organise cards into collections, export/import CSV
All card data lives in a single JSON file. No external API keys are required — you (the agent) generate flashcard content and quiz questions directly.
User-facing response style for Memento Flashcards:
- Use plain text only. Do not use Markdown formatting in replies to the user.
- Keep review and quiz feedback brief and neutral. Avoid extra praise, pep, or long explanations.
## When to Use
Use this skill when the user wants to:
- Save facts as flashcards for later review
- Review due cards with spaced repetition
- Generate a quiz from a YouTube video transcript
- Import, export, inspect, or delete flashcard data
Do not use this skill for general Q&A, coding help, or non-memory tasks.
## Quick Reference
| User intent | Action |
|---|---|
| "Remember that X" / "save this as a flashcard" | Generate a Q/A card, call `memento_cards.py add` |
| Sends a fact without mentioning flashcards | Ask "Want me to save this as a Memento flashcard?" — only create if confirmed |
| "Create a flashcard" | Ask for Q, A, collection; call `memento_cards.py add` |
| "Review my cards" | Call `memento_cards.py due`, present cards one-by-one |
| "Quiz me on [YouTube URL]" | Call `youtube_quiz.py fetch VIDEO_ID`, generate 5 questions, call `memento_cards.py add-quiz` |
| "Export my cards" | Call `memento_cards.py export --output PATH` |
| "Import cards from CSV" | Call `memento_cards.py import --file PATH --collection NAME` |
| "Show my stats" | Call `memento_cards.py stats` |
| "Delete a card" | Call `memento_cards.py delete --id ID` |
| "Delete a collection" | Call `memento_cards.py delete-collection --collection NAME` |
## Card Storage
Cards are stored in a JSON file at:
```
~/.hermes/skills/productivity/memento-flashcards/data/cards.json
```
**Never edit this file directly.** Always use `memento_cards.py` subcommands. The script handles atomic writes (write to temp file, then rename) to prevent corruption.
The file is created automatically on first use.
## Procedure
### Creating Cards from Facts
### Activation Rules
Not every factual statement should become a flashcard. Use this three-tier check:
1. **Explicit intent** — the user mentions "memento", "flashcard", "remember this", "save this card", "add a card", or similar phrasing that clearly requests a flashcard → **create the card directly**, no confirmation needed.
2. **Implicit intent** — the user sends a factual statement without mentioning flashcards (e.g. "The speed of light is 299,792 km/s") → **ask first**: "Want me to save this as a Memento flashcard?" Only create the card if the user confirms.
3. **No intent** — the message is a coding task, a question, instructions, normal conversation, or anything that is clearly not a fact to memorize → **do NOT activate this skill at all**. Let other skills or default behavior handle it.
When activation is confirmed (tier 1 directly, tier 2 after confirmation), generate a flashcard:
**Step 1:** Turn the statement into a Q/A pair. Use this format internally:
```
Turn the factual statement into a front-back pair.
Return exactly two lines:
Q: <question text>
A: <answer text>
Statement: "{statement}"
```
Rules:
- The question should test recall of the key fact
- The answer should be concise and direct
**Step 2:** Call the script to store the card:
```bash
python3 ~/.hermes/skills/productivity/memento-flashcards/scripts/memento_cards.py add \
--question "What year did World War 2 end?" \
--answer "1945" \
--collection "History"
```
If the user doesn't specify a collection, use `"General"` as the default.
The script outputs JSON confirming the created card.
### Manual Card Creation
When the user explicitly asks to create a flashcard, ask them for:
1. The question (front of card)
2. The answer (back of card)
3. The collection name (optional — default to `"General"`)
Then call `memento_cards.py add` as above.
### Reviewing Due Cards
When the user wants to review, fetch all due cards:
```bash
python3 ~/.hermes/skills/productivity/memento-flashcards/scripts/memento_cards.py due
```
This returns a JSON array of cards where `next_review_at <= now`. If a collection filter is needed:
```bash
python3 ~/.hermes/skills/productivity/memento-flashcards/scripts/memento_cards.py due --collection "History"
```
**Review flow (free-text grading):**
Here is an example of the EXACT interaction pattern you must follow. The user answers, you grade them, tell them the correct answer, then rate the card.
**Example interaction:**
> **Agent:** What year did the Berlin Wall fall?
>
> **User:** 1991
>
> **Agent:** Not quite. The Berlin Wall fell in 1989. Next review is tomorrow.
> *(agent calls: memento_cards.py rate --id ABC --rating hard --user-answer "1991")*
>
> Next question: Who was the first person to walk on the moon?
**The rules:**
1. Show only the question. Wait for the user to answer.
2. After receiving their answer, compare it to the expected answer and grade it:
- **correct** → user got the key fact right (even if worded differently)
- **partial** → right track but missing the core detail
- **incorrect** → wrong or off-topic
3. **You MUST tell the user the correct answer and how they did.** Keep it short and plain-text. Use this format:
- correct: "Correct. Answer: &#123;answer&#125;. Next review in 7 days."
- partial: "Close. Answer: &#123;answer&#125;. &#123;what they missed&#125;. Next review in 3 days."
- incorrect: "Not quite. Answer: &#123;answer&#125;. Next review tomorrow."
4. Then call the rate command: correct→easy, partial→good, incorrect→hard.
5. Then show the next question.
```bash
python3 ~/.hermes/skills/productivity/memento-flashcards/scripts/memento_cards.py rate \
--id CARD_ID --rating easy --user-answer "what the user said"
```
**Never skip step 3.** The user must always see the correct answer and feedback before you move on.
If no cards are due, tell the user: "No cards due for review right now. Check back later!"
**Retire override:** At any point the user can say "retire this card" to permanently remove it from reviews. Use `--rating retire` for this.
### Spaced Repetition Algorithm
The rating determines the next review interval:
| Rating | Interval | ease_streak | Status change |
|---|---|---|---|
| **hard** | +1 day | reset to 0 | stays learning |
| **good** | +3 days | reset to 0 | stays learning |
| **easy** | +7 days | +1 | if ease_streak >= 3 → retired |
| **retire** | permanent | reset to 0 | → retired |
- **learning**: card is actively in rotation
- **retired**: card won't appear in reviews (user has mastered it or manually retired it)
- Three consecutive "easy" ratings automatically retire a card
### YouTube Quiz Generation
When the user sends a YouTube URL and wants a quiz:
**Step 1:** Extract the video ID from the URL (e.g. `dQw4w9WgXcQ` from `https://www.youtube.com/watch?v=dQw4w9WgXcQ`).
**Step 2:** Fetch the transcript:
```bash
python3 ~/.hermes/skills/productivity/memento-flashcards/scripts/youtube_quiz.py fetch VIDEO_ID
```
This returns `{"title": "...", "transcript": "..."}` or an error.
If the script reports `missing_dependency`, tell the user to install it:
```bash
pip install youtube-transcript-api
```
**Step 3:** Generate 5 quiz questions from the transcript. Use these rules:
```
You are creating a 5-question quiz for a podcast episode.
Return ONLY a JSON array with exactly 5 objects.
Each object must contain keys 'question' and 'answer'.
Selection criteria:
- Prioritize important, surprising, or foundational facts.
- Skip filler, obvious details, and facts that require heavy context.
- Never return true/false questions.
- Never ask only for a date.
Question rules:
- Each question must test exactly one discrete fact.
- Use clear, unambiguous wording.
- Prefer What, Who, How many, Which.
- Avoid open-ended Describe or Explain prompts.
Answer rules:
- Each answer must be under 240 characters.
- Lead with the answer itself, not preamble.
- Add only minimal clarifying detail if needed.
```
Use the first 15,000 characters of the transcript as context. Generate the questions yourself (you are the LLM).
**Step 4:** Validate the output is valid JSON with exactly 5 items, each having non-empty `question` and `answer` strings. If validation fails, retry once.
**Step 5:** Store quiz cards:
```bash
python3 ~/.hermes/skills/productivity/memento-flashcards/scripts/memento_cards.py add-quiz \
--video-id "VIDEO_ID" \
--questions '[{"question":"...","answer":"..."},...]' \
--collection "Quiz - Episode Title"
```
The script deduplicates by `video_id` — if cards for that video already exist, it skips creation and reports the existing cards.
**Step 6:** Present questions one-by-one using the same free-text grading flow:
1. Show "Question 1/5: ..." and wait for the user's answer. Never include the answer or any hint about revealing it.
2. Wait for the user to answer in their own words
3. Grade their answer using the grading prompt (see "Reviewing Due Cards" section)
4. **IMPORTANT: You MUST reply to the user with feedback before doing anything else.** Show the grade, the correct answer, and when the card is next due. Do NOT silently skip to the next question. Keep it short and plain-text. Example: "Not quite. Answer: &#123;answer&#125;. Next review tomorrow."
5. **After showing feedback**, call the rate command and then show the next question in the same message:
```bash
python3 ~/.hermes/skills/productivity/memento-flashcards/scripts/memento_cards.py rate \
--id CARD_ID --rating easy --user-answer "what the user said"
```
6. Repeat. Every answer MUST receive visible feedback before the next question.
### Export/Import CSV
**Export:**
```bash
python3 ~/.hermes/skills/productivity/memento-flashcards/scripts/memento_cards.py export \
--output ~/flashcards.csv
```
Produces a 3-column CSV: `question,answer,collection` (no header row).
**Import:**
```bash
python3 ~/.hermes/skills/productivity/memento-flashcards/scripts/memento_cards.py import \
--file ~/flashcards.csv \
--collection "Imported"
```
Reads a CSV with columns: question, answer, and optionally collection (column 3). If the collection column is missing, uses the `--collection` argument.
### Statistics
```bash
python3 ~/.hermes/skills/productivity/memento-flashcards/scripts/memento_cards.py stats
```
Returns JSON with:
- `total`: total card count
- `learning`: cards in active rotation
- `retired`: mastered cards
- `due_now`: cards due for review right now
- `collections`: breakdown by collection name
## Pitfalls
- **Never edit `cards.json` directly** — always use the script subcommands to avoid corruption
- **Transcript failures** — some YouTube videos have no English transcript or have transcripts disabled; inform the user and suggest another video
- **Optional dependency** — `youtube_quiz.py` needs `youtube-transcript-api`; if missing, tell the user to run `pip install youtube-transcript-api`
- **Large imports** — CSV imports with thousands of rows work fine but the JSON output may be verbose; summarize the result for the user
- **Video ID extraction** — support both `youtube.com/watch?v=ID` and `youtu.be/ID` URL formats
## Verification
Verify the helper scripts directly:
```bash
python3 ~/.hermes/skills/productivity/memento-flashcards/scripts/memento_cards.py stats
python3 ~/.hermes/skills/productivity/memento-flashcards/scripts/memento_cards.py add --question "Capital of France?" --answer "Paris" --collection "General"
python3 ~/.hermes/skills/productivity/memento-flashcards/scripts/memento_cards.py due
```
If you are testing from the repo checkout, run:
```bash
pytest tests/skills/test_memento_cards.py tests/skills/test_youtube_quiz.py -q
```
Agent-level verification:
- Start a review and confirm feedback is plain text, brief, and always includes the correct answer before the next card
- Run a YouTube quiz flow and confirm each answer receives visible feedback before the next question

304
website/docs/user-guide/skills/optional/productivity/productivity-siyuan.md Normal file
View file

@ -0,0 +1,304 @@
---
title: "Siyuan"
sidebar_label: "Siyuan"
description: "SiYuan Note API for searching, reading, creating, and managing blocks and documents in a self-hosted knowledge base via curl"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Siyuan
SiYuan Note API for searching, reading, creating, and managing blocks and documents in a self-hosted knowledge base via curl.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/productivity/siyuan` |
| Path | `optional-skills/productivity/siyuan` |
| Version | `1.0.0` |
| Author | FEUAZUR |
| License | MIT |
| Tags | `SiYuan`, `Notes`, `Knowledge Base`, `PKM`, `API` |
| Related skills | [`obsidian`](/docs/user-guide/skills/bundled/note-taking/note-taking-obsidian), [`notion`](/docs/user-guide/skills/bundled/productivity/productivity-notion) |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# SiYuan Note API
Use the [SiYuan](https://github.com/siyuan-note/siyuan) kernel API via curl to search, read, create, update, and delete blocks and documents in a self-hosted knowledge base. No extra tools needed -- just curl and an API token.
## Prerequisites
1. Install and run SiYuan (desktop or Docker)
2. Get your API token: **Settings > About > API token**
3. Store it in `~/.hermes/.env`:
```
SIYUAN_TOKEN=your_token_here
SIYUAN_URL=http://127.0.0.1:6806
```
`SIYUAN_URL` defaults to `http://127.0.0.1:6806` if not set.
## API Basics
All SiYuan API calls are **POST with JSON body**. Every request follows this pattern:
```bash
curl -s -X POST "${SIYUAN_URL:-http://127.0.0.1:6806}/api/..." \
-H "Authorization: Token $SIYUAN_TOKEN" \
-H "Content-Type: application/json" \
-d '{"param": "value"}'
```
Responses are JSON with this structure:
```json
{"code": 0, "msg": "", "data": { ... }}
```
`code: 0` means success. Any other value is an error -- check `msg` for details.
**ID format:** SiYuan IDs look like `20210808180117-6v0mkxr` (14-digit timestamp + 7 alphanumeric chars).
## Quick Reference
| Operation | Endpoint |
|-----------|----------|
| Full-text search | `/api/search/fullTextSearchBlock` |
| SQL query | `/api/query/sql` |
| Read block | `/api/block/getBlockKramdown` |
| Read children | `/api/block/getChildBlocks` |
| Get path | `/api/filetree/getHPathByID` |
| Get attributes | `/api/attr/getBlockAttrs` |
| List notebooks | `/api/notebook/lsNotebooks` |
| List documents | `/api/filetree/listDocsByPath` |
| Create notebook | `/api/notebook/createNotebook` |
| Create document | `/api/filetree/createDocWithMd` |
| Append block | `/api/block/appendBlock` |
| Update block | `/api/block/updateBlock` |
| Rename document | `/api/filetree/renameDocByID` |
| Set attributes | `/api/attr/setBlockAttrs` |
| Delete block | `/api/block/deleteBlock` |
| Delete document | `/api/filetree/removeDocByID` |
| Export as Markdown | `/api/export/exportMdContent` |
## Common Operations
### Search (Full-Text)
```bash
curl -s -X POST "${SIYUAN_URL:-http://127.0.0.1:6806}/api/search/fullTextSearchBlock" \
-H "Authorization: Token $SIYUAN_TOKEN" \
-H "Content-Type: application/json" \
-d '{"query": "meeting notes", "page": 0}' | jq '.data.blocks[:5]'
```
### Search (SQL)
Query the blocks database directly. Only SELECT statements are safe.
```bash
curl -s -X POST "${SIYUAN_URL:-http://127.0.0.1:6806}/api/query/sql" \
-H "Authorization: Token $SIYUAN_TOKEN" \
-H "Content-Type: application/json" \
-d '{"stmt": "SELECT id, content, type, box FROM blocks WHERE content LIKE '\''%keyword%'\'' AND type='\''p'\'' LIMIT 20"}' | jq '.data'
```
Useful columns: `id`, `parent_id`, `root_id`, `box` (notebook ID), `path`, `content`, `type`, `subtype`, `created`, `updated`.
### Read Block Content
Returns block content in Kramdown (Markdown-like) format.
```bash
curl -s -X POST "${SIYUAN_URL:-http://127.0.0.1:6806}/api/block/getBlockKramdown" \
-H "Authorization: Token $SIYUAN_TOKEN" \
-H "Content-Type: application/json" \
-d '{"id": "20210808180117-6v0mkxr"}' | jq '.data.kramdown'
```
### Read Child Blocks
```bash
curl -s -X POST "${SIYUAN_URL:-http://127.0.0.1:6806}/api/block/getChildBlocks" \
-H "Authorization: Token $SIYUAN_TOKEN" \
-H "Content-Type: application/json" \
-d '{"id": "20210808180117-6v0mkxr"}' | jq '.data'
```
### Get Human-Readable Path
```bash
curl -s -X POST "${SIYUAN_URL:-http://127.0.0.1:6806}/api/filetree/getHPathByID" \
-H "Authorization: Token $SIYUAN_TOKEN" \
-H "Content-Type: application/json" \
-d '{"id": "20210808180117-6v0mkxr"}' | jq '.data'
```
### Get Block Attributes
```bash
curl -s -X POST "${SIYUAN_URL:-http://127.0.0.1:6806}/api/attr/getBlockAttrs" \
-H "Authorization: Token $SIYUAN_TOKEN" \
-H "Content-Type: application/json" \
-d '{"id": "20210808180117-6v0mkxr"}' | jq '.data'
```
### List Notebooks
```bash
curl -s -X POST "${SIYUAN_URL:-http://127.0.0.1:6806}/api/notebook/lsNotebooks" \
-H "Authorization: Token $SIYUAN_TOKEN" \
-H "Content-Type: application/json" \
-d '{}' | jq '.data.notebooks[] | {id, name, closed}'
```
### List Documents in a Notebook
```bash
curl -s -X POST "${SIYUAN_URL:-http://127.0.0.1:6806}/api/filetree/listDocsByPath" \
-H "Authorization: Token $SIYUAN_TOKEN" \
-H "Content-Type: application/json" \
-d '{"notebook": "NOTEBOOK_ID", "path": "/"}' | jq '.data.files[] | {id, name}'
```
### Create a Document
```bash
curl -s -X POST "${SIYUAN_URL:-http://127.0.0.1:6806}/api/filetree/createDocWithMd" \
-H "Authorization: Token $SIYUAN_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"notebook": "NOTEBOOK_ID",
"path": "/Meeting Notes/2026-03-22",
"markdown": "# Meeting Notes\n\n- Discussed project timeline\n- Assigned tasks"
}' | jq '.data'
```
### Create a Notebook
```bash
curl -s -X POST "${SIYUAN_URL:-http://127.0.0.1:6806}/api/notebook/createNotebook" \
-H "Authorization: Token $SIYUAN_TOKEN" \
-H "Content-Type: application/json" \
-d '{"name": "My New Notebook"}' | jq '.data.notebook.id'
```
### Append Block to Document
```bash
curl -s -X POST "${SIYUAN_URL:-http://127.0.0.1:6806}/api/block/appendBlock" \
-H "Authorization: Token $SIYUAN_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"parentID": "DOCUMENT_OR_BLOCK_ID",
"data": "New paragraph added at the end.",
"dataType": "markdown"
}' | jq '.data'
```
Also available: `/api/block/prependBlock` (same params, inserts at the beginning) and `/api/block/insertBlock` (uses `previousID` instead of `parentID` to insert after a specific block).
### Update Block Content
```bash
curl -s -X POST "${SIYUAN_URL:-http://127.0.0.1:6806}/api/block/updateBlock" \
-H "Authorization: Token $SIYUAN_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"id": "BLOCK_ID",
"data": "Updated content here.",
"dataType": "markdown"
}' | jq '.data'
```
### Rename a Document
```bash
curl -s -X POST "${SIYUAN_URL:-http://127.0.0.1:6806}/api/filetree/renameDocByID" \
-H "Authorization: Token $SIYUAN_TOKEN" \
-H "Content-Type: application/json" \
-d '{"id": "DOCUMENT_ID", "title": "New Title"}'
```
### Set Block Attributes
Custom attributes must be prefixed with `custom-`:
```bash
curl -s -X POST "${SIYUAN_URL:-http://127.0.0.1:6806}/api/attr/setBlockAttrs" \
-H "Authorization: Token $SIYUAN_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"id": "BLOCK_ID",
"attrs": {
"custom-status": "reviewed",
"custom-priority": "high"
}
}'
```
### Delete a Block
```bash
curl -s -X POST "${SIYUAN_URL:-http://127.0.0.1:6806}/api/block/deleteBlock" \
-H "Authorization: Token $SIYUAN_TOKEN" \
-H "Content-Type: application/json" \
-d '{"id": "BLOCK_ID"}'
```
To delete a whole document: use `/api/filetree/removeDocByID` with `{"id": "DOC_ID"}`.
To delete a notebook: use `/api/notebook/removeNotebook` with `{"notebook": "NOTEBOOK_ID"}`.
### Export Document as Markdown
```bash
curl -s -X POST "${SIYUAN_URL:-http://127.0.0.1:6806}/api/export/exportMdContent" \
-H "Authorization: Token $SIYUAN_TOKEN" \
-H "Content-Type: application/json" \
-d '{"id": "DOCUMENT_ID"}' | jq -r '.data.content'
```
## Block Types
Common `type` values in SQL queries:
| Type | Description |
|------|-------------|
| `d` | Document (root block) |
| `p` | Paragraph |
| `h` | Heading |
| `l` | List |
| `i` | List item |
| `c` | Code block |
| `m` | Math block |
| `t` | Table |
| `b` | Blockquote |
| `s` | Super block |
| `html` | HTML block |
## Pitfalls
- **All endpoints are POST** -- even read-only operations. Do not use GET.
- **SQL safety**: only use SELECT queries. INSERT/UPDATE/DELETE/DROP are dangerous and should never be sent.
- **ID validation**: IDs match the pattern `YYYYMMDDHHmmss-xxxxxxx`. Reject anything else.
- **Error responses**: always check `code != 0` in responses before processing `data`.
- **Large documents**: block content and export results can be very large. Use `LIMIT` in SQL and pipe through `jq` to extract only what you need.
- **Notebook IDs**: when working with a specific notebook, get its ID first via `lsNotebooks`.
## Alternative: MCP Server
If you prefer a native integration instead of curl, install the SiYuan MCP server:
```yaml
# In ~/.hermes/config.yaml under mcp_servers:
mcp_servers:
siyuan:
command: npx
args: ["-y", "@porkll/siyuan-mcp"]
env:
SIYUAN_TOKEN: "your_token"
SIYUAN_URL: "http://127.0.0.1:6806"
```

434
website/docs/user-guide/skills/optional/productivity/productivity-telephony.md Normal file
View file

@ -0,0 +1,434 @@
---
title: "Telephony — Give Hermes phone capabilities without core tool changes"
sidebar_label: "Telephony"
description: "Give Hermes phone capabilities without core tool changes"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Telephony
Give Hermes phone capabilities without core tool changes. Provision and persist a Twilio number, send and receive SMS/MMS, make direct calls, and place AI-driven outbound calls through Bland.ai or Vapi.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/productivity/telephony` |
| Path | `optional-skills/productivity/telephony` |
| Version | `1.0.0` |
| Author | Nous Research |
| License | MIT |
| Tags | `telephony`, `phone`, `sms`, `mms`, `voice`, `twilio`, `bland.ai`, `vapi`, `calling`, `texting` |
| Related skills | [`maps`](/docs/user-guide/skills/bundled/productivity/productivity-maps), [`google-workspace`](/docs/user-guide/skills/bundled/productivity/productivity-google-workspace), [`agentmail`](/docs/user-guide/skills/optional/email/email-agentmail) |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Telephony — Numbers, Calls, and Texts without Core Tool Changes
This optional skill gives Hermes practical phone capabilities while keeping telephony out of the core tool list.
It ships with a helper script, `scripts/telephony.py`, that can:
- save provider credentials into `~/.hermes/.env`
- search for and buy a Twilio phone number
- remember that owned number for later sessions
- send SMS / MMS from the owned number
- poll inbound SMS for that number with no webhook server required
- make direct Twilio calls using TwiML `<Say>` or `<Play>`
- import the owned Twilio number into Vapi
- place outbound AI calls through Bland.ai or Vapi
## What this solves
This skill is meant to cover the practical phone tasks users actually want:
- outbound calls
- texting
- owning a reusable agent number
- checking messages that arrive to that number later
- preserving that number and related IDs between sessions
- future-friendly telephony identity for inbound SMS polling and other automations
It does **not** turn Hermes into a real-time inbound phone gateway. Inbound SMS is handled by polling the Twilio REST API. That is enough for many workflows, including notifications and some one-time-code retrieval, without adding core webhook infrastructure.
## Safety rules — mandatory
1. Always confirm before placing a call or sending a text.
2. Never dial emergency numbers.
3. Never use telephony for harassment, spam, impersonation, or anything illegal.
4. Treat third-party phone numbers as sensitive operational data:
- do not save them to Hermes memory
- do not include them in skill docs, summaries, or follow-up notes unless the user explicitly wants that
5. It is fine to persist the **agent-owned Twilio number** because that is part of the user's configuration.
6. VoIP numbers are **not guaranteed** to work for all third-party 2FA flows. Use with caution and set user expectations clearly.
## Decision tree — which service to use?
Use this logic instead of hardcoded provider routing:
### 1) "I want Hermes to own a real phone number"
Use **Twilio**.
Why:
- easiest path to buying and keeping a number
- best SMS / MMS support
- simplest inbound SMS polling story
- cleanest future path to inbound webhooks or call handling
Use cases:
- receive texts later
- send deployment alerts / cron notifications
- maintain a reusable phone identity for the agent
- experiment with phone-based auth flows later
### 2) "I only need the easiest outbound AI phone call right now"
Use **Bland.ai**.
Why:
- quickest setup
- one API key
- no need to first buy/import a number yourself
Tradeoff:
- less flexible
- voice quality is decent, but not the best
### 3) "I want the best conversational AI voice quality"
Use **Twilio + Vapi**.
Why:
- Twilio gives you the owned number
- Vapi gives you better conversational AI call quality and more voice/model flexibility
Recommended flow:
1. Buy/save a Twilio number
2. Import it into Vapi
3. Save the returned `VAPI_PHONE_NUMBER_ID`
4. Use `ai-call --provider vapi`
### 4) "I want to call with a custom prerecorded voice message"
Use **Twilio direct call** with a public audio URL.
Why:
- easiest way to play a custom MP3
- pairs well with Hermes `text_to_speech` plus a public file host or tunnel
## Files and persistent state
The skill persists telephony state in two places:
### `~/.hermes/.env`
Used for long-lived provider credentials and owned-number IDs, for example:
- `TWILIO_ACCOUNT_SID`
- `TWILIO_AUTH_TOKEN`
- `TWILIO_PHONE_NUMBER`
- `TWILIO_PHONE_NUMBER_SID`
- `BLAND_API_KEY`
- `VAPI_API_KEY`
- `VAPI_PHONE_NUMBER_ID`
- `PHONE_PROVIDER` (AI call provider: bland or vapi)
### `~/.hermes/telephony_state.json`
Used for skill-only state that should survive across sessions, for example:
- remembered default Twilio number / SID
- remembered Vapi phone number ID
- last inbound message SID/date for inbox polling checkpoints
This means:
- the next time the skill is loaded, `diagnose` can tell you what number is already configured
- `twilio-inbox --since-last --mark-seen` can continue from the previous checkpoint
## Locate the helper script
After installing this skill, locate the script like this:
```bash
SCRIPT="$(find ~/.hermes/skills -path '*/telephony/scripts/telephony.py' -print -quit)"
```
If `SCRIPT` is empty, the skill is not installed yet.
## Install
This is an official optional skill, so install it from the Skills Hub:
```bash
hermes skills search telephony
hermes skills install official/productivity/telephony
```
## Provider setup
### Twilio — owned number, SMS/MMS, direct calls, inbound SMS polling
Sign up at:
- https://www.twilio.com/try-twilio
Then save credentials into Hermes:
```bash
python3 "$SCRIPT" save-twilio ACXXXXXXXXXXXXXXXXXXXXXXXXXXXX your_auth_token_here
```
Search for available numbers:
```bash
python3 "$SCRIPT" twilio-search --country US --area-code 702 --limit 5
```
Buy and remember a number:
```bash
python3 "$SCRIPT" twilio-buy "+17025551234" --save-env
```
List owned numbers:
```bash
python3 "$SCRIPT" twilio-owned
```
Set one of them as the default later:
```bash
python3 "$SCRIPT" twilio-set-default "+17025551234" --save-env
# or
python3 "$SCRIPT" twilio-set-default PNXXXXXXXXXXXXXXXXXXXXXXXXXXXX --save-env
```
### Bland.ai — easiest outbound AI calling
Sign up at:
- https://app.bland.ai
Save config:
```bash
python3 "$SCRIPT" save-bland your_bland_api_key --voice mason
```
### Vapi — better conversational voice quality
Sign up at:
- https://dashboard.vapi.ai
Save the API key first:
```bash
python3 "$SCRIPT" save-vapi your_vapi_api_key
```
Import your owned Twilio number into Vapi and persist the returned phone number ID:
```bash
python3 "$SCRIPT" vapi-import-twilio --save-env
```
If you already know the Vapi phone number ID, save it directly:
```bash
python3 "$SCRIPT" save-vapi your_vapi_api_key --phone-number-id vapi_phone_number_id_here
```
## Diagnose current state
At any time, inspect what the skill already knows:
```bash
python3 "$SCRIPT" diagnose
```
Use this first when resuming work in a later session.
## Common workflows
### A. Buy an agent number and keep using it later
1. Save Twilio credentials:
```bash
python3 "$SCRIPT" save-twilio AC... auth_token_here
```
2. Search for a number:
```bash
python3 "$SCRIPT" twilio-search --country US --area-code 702 --limit 10
```
3. Buy it and save it into `~/.hermes/.env` + state:
```bash
python3 "$SCRIPT" twilio-buy "+17025551234" --save-env
```
4. Next session, run:
```bash
python3 "$SCRIPT" diagnose
```
This shows the remembered default number and inbox checkpoint state.
### B. Send a text from the agent number
```bash
python3 "$SCRIPT" twilio-send-sms "+15551230000" "Your deployment completed successfully."
```
With media:
```bash
python3 "$SCRIPT" twilio-send-sms "+15551230000" "Here is the chart." --media-url "https://example.com/chart.png"
```
### C. Check inbound texts later with no webhook server
Poll the inbox for the default Twilio number:
```bash
python3 "$SCRIPT" twilio-inbox --limit 20
```
Only show messages that arrived after the last checkpoint, and advance the checkpoint when you're done reading:
```bash
python3 "$SCRIPT" twilio-inbox --since-last --mark-seen
```
This is the main answer to “how do I access messages the number receives next time the skill is loaded?”
### D. Make a direct Twilio call with built-in TTS
```bash
python3 "$SCRIPT" twilio-call "+15551230000" --message "Hello! This is Hermes calling with your status update." --voice Polly.Joanna
```
### E. Call with a prerecorded / custom voice message
This is the main path for reusing Hermes's existing `text_to_speech` support.
Use this when:
- you want the call to use Hermes's configured TTS voice rather than Twilio `<Say>`
- you want a one-way voice delivery (briefing, alert, joke, reminder, status update)
- you do **not** need a live conversational phone call
Generate or host audio separately, then:
```bash
python3 "$SCRIPT" twilio-call "+155****0000" --audio-url "https://example.com/briefing.mp3"
```
Recommended Hermes TTS -> Twilio Play workflow:
1. Generate the audio with Hermes `text_to_speech`.
2. Make the resulting MP3 publicly reachable.
3. Place the Twilio call with `--audio-url`.
Example agent flow:
- Ask Hermes to create the message audio with `text_to_speech`
- If needed, expose the file with a temporary static host / tunnel / object storage URL
- Use `twilio-call --audio-url ...` to deliver it by phone
Good hosting options for the MP3:
- a temporary public object/storage URL
- a short-lived tunnel to a local static file server
- any existing HTTPS URL the phone provider can fetch directly
Important note:
- Hermes TTS is great for prerecorded outbound messages
- Bland/Vapi are better for **live conversational AI calls** because they handle the real-time telephony audio stack themselves
- Hermes STT/TTS alone is not being used here as a full duplex phone conversation engine; that would require a much heavier streaming/webhook integration than this skill is trying to introduce
### F. Navigate a phone tree / IVR with Twilio direct calling
If you need to press digits after the call connects, use `--send-digits`.
Twilio interprets `w` as a short wait.
```bash
python3 "$SCRIPT" twilio-call "+18005551234" --message "Connecting to billing now." --send-digits "ww1w2w3"
```
This is useful for reaching a specific menu branch before handing off to a human or delivering a short status message.
### G. Outbound AI phone call with Bland.ai
```bash
python3 "$SCRIPT" ai-call "+15551230000" "Call the dental office, ask for a cleaning appointment on Tuesday afternoon, and if they do not have Tuesday availability, ask for Wednesday or Thursday instead." --provider bland --voice mason --max-duration 3
```
Check status:
```bash
python3 "$SCRIPT" ai-status <call_id> --provider bland
```
Ask Bland analysis questions after completion:
```bash
python3 "$SCRIPT" ai-status <call_id> --provider bland --analyze "Was the appointment confirmed?,What date and time?,Any special instructions?"
```
### H. Outbound AI phone call with Vapi on your owned number
1. Import your Twilio number into Vapi:
```bash
python3 "$SCRIPT" vapi-import-twilio --save-env
```
2. Place the call:
```bash
python3 "$SCRIPT" ai-call "+15551230000" "You are calling to make a dinner reservation for two at 7:30 PM. If that is unavailable, ask for the nearest time between 6:30 and 8:30 PM." --provider vapi --max-duration 4
```
3. Check result:
```bash
python3 "$SCRIPT" ai-status <call_id> --provider vapi
```
## Suggested agent procedure
When the user asks for a call or text:
1. Determine which path fits the request via the decision tree.
2. Run `diagnose` if configuration state is unclear.
3. Gather the full task details.
4. Confirm with the user before dialing or texting.
5. Use the correct command.
6. Poll for results if needed.
7. Summarize the outcome without persisting third-party numbers to Hermes memory.
## What this skill still does not do
- real-time inbound call answering
- webhook-based live SMS push into the agent loop
- guaranteed support for arbitrary third-party 2FA providers
Those would require more infrastructure than a pure optional skill.
## Pitfalls
- Twilio trial accounts and regional rules can restrict who you can call/text.
- Some services reject VoIP numbers for 2FA.
- `twilio-inbox` polls the REST API; it is not instant push delivery.
- Vapi outbound calling still depends on having a valid imported number.
- Bland is easiest, but not always the best-sounding.
- Do not store arbitrary third-party phone numbers in Hermes memory.
## Verification checklist
After setup, you should be able to do all of the following with just this skill:
1. `diagnose` shows provider readiness and remembered state
2. search and buy a Twilio number
3. persist that number to `~/.hermes/.env`
4. send an SMS from the owned number
5. poll inbound texts for the owned number later
6. place a direct Twilio call
7. place an AI call via Bland or Vapi
## References
- Twilio phone numbers: https://www.twilio.com/docs/phone-numbers/api
- Twilio messaging: https://www.twilio.com/docs/messaging/api/message-resource
- Twilio voice: https://www.twilio.com/docs/voice/api/call-resource
- Vapi docs: https://docs.vapi.ai/
- Bland.ai: https://app.bland.ai/

252
website/docs/user-guide/skills/optional/research/research-bioinformatics.md Normal file
View file

@ -0,0 +1,252 @@
---
title: "Bioinformatics — Gateway to 400+ bioinformatics skills from bioSkills and ClawBio"
sidebar_label: "Bioinformatics"
description: "Gateway to 400+ bioinformatics skills from bioSkills and ClawBio"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Bioinformatics
Gateway to 400+ bioinformatics skills from bioSkills and ClawBio. Covers genomics, transcriptomics, single-cell, variant calling, pharmacogenomics, metagenomics, structural biology, and more. Fetches domain-specific reference material on demand.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/research/bioinformatics` |
| Path | `optional-skills/research/bioinformatics` |
| Version | `1.0.0` |
| Platforms | linux, macos |
| Tags | `bioinformatics`, `genomics`, `sequencing`, `biology`, `research`, `science` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Bioinformatics Skills Gateway
Use when asked about bioinformatics, genomics, sequencing, variant calling, gene expression, single-cell analysis, protein structure, pharmacogenomics, metagenomics, phylogenetics, or any computational biology task.
This skill is a gateway to two open-source bioinformatics skill libraries. Instead of bundling hundreds of domain-specific skills, it indexes them and fetches what you need on demand.
## Sources
◆ **bioSkills** — 385 reference skills (code patterns, parameter guides, decision trees)
Repo: https://github.com/GPTomics/bioSkills
Format: SKILL.md per topic with code examples. Python/R/CLI.
◆ **ClawBio** — 33 runnable pipeline skills (executable scripts, reproducibility bundles)
Repo: https://github.com/ClawBio/ClawBio
Format: Python scripts with demos. Each analysis exports report.md + commands.sh + environment.yml.
## How to fetch and use a skill
1. Identify the domain and skill name from the index below.
2. Clone the relevant repo (shallow clone to save time):
```bash
# bioSkills (reference material)
git clone --depth 1 https://github.com/GPTomics/bioSkills.git /tmp/bioSkills
# ClawBio (runnable pipelines)
git clone --depth 1 https://github.com/ClawBio/ClawBio.git /tmp/ClawBio
```
3. Read the specific skill:
```bash
# bioSkills — each skill is at: <category>/<skill-name>/SKILL.md
cat /tmp/bioSkills/variant-calling/gatk-variant-calling/SKILL.md
# ClawBio — each skill is at: skills/<skill-name>/
cat /tmp/ClawBio/skills/pharmgx-reporter/README.md
```
4. Follow the fetched skill as reference material. These are NOT Hermes-format skills — treat them as expert domain guides. They contain correct parameters, proper tool flags, and validated pipelines.
## Skill Index by Domain
### Sequence Fundamentals
bioSkills:
sequence-io/ — read-sequences, write-sequences, format-conversion, batch-processing, compressed-files, fastq-quality, filter-sequences, paired-end-fastq, sequence-statistics
sequence-manipulation/ — seq-objects, reverse-complement, transcription-translation, motif-search, codon-usage, sequence-properties, sequence-slicing
ClawBio:
seq-wrangler — Sequence QC, alignment, and BAM processing (wraps FastQC, BWA, SAMtools)
### Read QC & Alignment
bioSkills:
read-qc/ — quality-reports, fastp-workflow, adapter-trimming, quality-filtering, umi-processing, contamination-screening, rnaseq-qc
read-alignment/ — bwa-alignment, star-alignment, hisat2-alignment, bowtie2-alignment
alignment-files/ — sam-bam-basics, alignment-sorting, alignment-filtering, bam-statistics, duplicate-handling, pileup-generation
### Variant Calling & Annotation
bioSkills:
variant-calling/ — gatk-variant-calling, deepvariant, variant-calling (bcftools), joint-calling, structural-variant-calling, filtering-best-practices, variant-annotation, variant-normalization, vcf-basics, vcf-manipulation, vcf-statistics, consensus-sequences, clinical-interpretation
ClawBio:
vcf-annotator — VEP + ClinVar + gnomAD annotation with ancestry-aware context
variant-annotation — Variant annotation pipeline
### Differential Expression (Bulk RNA-seq)
bioSkills:
differential-expression/ — deseq2-basics, edger-basics, batch-correction, de-results, de-visualization, timeseries-de
rna-quantification/ — alignment-free-quant (Salmon/kallisto), featurecounts-counting, tximport-workflow, count-matrix-qc
expression-matrix/ — counts-ingest, gene-id-mapping, metadata-joins, sparse-handling
ClawBio:
rnaseq-de — Full DE pipeline with QC, normalization, and visualization
diff-visualizer — Rich visualization and reporting for DE results
### Single-Cell RNA-seq
bioSkills:
single-cell/ — preprocessing, clustering, batch-integration, cell-annotation, cell-communication, doublet-detection, markers-annotation, trajectory-inference, multimodal-integration, perturb-seq, scatac-analysis, lineage-tracing, metabolite-communication, data-io
ClawBio:
scrna-orchestrator — Full Scanpy pipeline (QC, clustering, markers, annotation)
scrna-embedding — scVI-based latent embedding and batch integration
### Spatial Transcriptomics
bioSkills:
spatial-transcriptomics/ — spatial-data-io, spatial-preprocessing, spatial-domains, spatial-deconvolution, spatial-communication, spatial-neighbors, spatial-statistics, spatial-visualization, spatial-multiomics, spatial-proteomics, image-analysis
### Epigenomics
bioSkills:
chip-seq/ — peak-calling, differential-binding, motif-analysis, peak-annotation, chipseq-qc, chipseq-visualization, super-enhancers
atac-seq/ — atac-peak-calling, atac-qc, differential-accessibility, footprinting, motif-deviation, nucleosome-positioning
methylation-analysis/ — bismark-alignment, methylation-calling, dmr-detection, methylkit-analysis
hi-c-analysis/ — hic-data-io, tad-detection, loop-calling, compartment-analysis, contact-pairs, matrix-operations, hic-visualization, hic-differential
ClawBio:
methylation-clock — Epigenetic age estimation
### Pharmacogenomics & Clinical
bioSkills:
clinical-databases/ — clinvar-lookup, gnomad-frequencies, dbsnp-queries, pharmacogenomics, polygenic-risk, hla-typing, variant-prioritization, somatic-signatures, tumor-mutational-burden, myvariant-queries
ClawBio:
pharmgx-reporter — PGx report from 23andMe/AncestryDNA (12 genes, 31 SNPs, 51 drugs)
drug-photo — Photo of medication → personalized PGx dosage card (via vision)
clinpgx — ClinPGx API for gene-drug data and CPIC guidelines
gwas-lookup — Federated variant lookup across 9 genomic databases
gwas-prs — Polygenic risk scores from consumer genetic data
nutrigx_advisor — Personalized nutrition from consumer genetic data
### Population Genetics & GWAS
bioSkills:
population-genetics/ — association-testing (PLINK GWAS), plink-basics, population-structure, linkage-disequilibrium, scikit-allel-analysis, selection-statistics
causal-genomics/ — mendelian-randomization, fine-mapping, colocalization-analysis, mediation-analysis, pleiotropy-detection
phasing-imputation/ — haplotype-phasing, genotype-imputation, imputation-qc, reference-panels
ClawBio:
claw-ancestry-pca — Ancestry PCA against SGDP reference panel
### Metagenomics & Microbiome
bioSkills:
metagenomics/ — kraken-classification, metaphlan-profiling, abundance-estimation, functional-profiling, amr-detection, strain-tracking, metagenome-visualization
microbiome/ — amplicon-processing, diversity-analysis, differential-abundance, taxonomy-assignment, functional-prediction, qiime2-workflow
ClawBio:
claw-metagenomics — Shotgun metagenomics profiling (taxonomy, resistome, functional pathways)
### Genome Assembly & Annotation
bioSkills:
genome-assembly/ — hifi-assembly, long-read-assembly, short-read-assembly, metagenome-assembly, assembly-polishing, assembly-qc, scaffolding, contamination-detection
genome-annotation/ — eukaryotic-gene-prediction, prokaryotic-annotation, functional-annotation, ncrna-annotation, repeat-annotation, annotation-transfer
long-read-sequencing/ — basecalling, long-read-alignment, long-read-qc, clair3-variants, structural-variants, medaka-polishing, nanopore-methylation, isoseq-analysis
### Structural Biology & Chemoinformatics
bioSkills:
structural-biology/ — alphafold-predictions, modern-structure-prediction, structure-io, structure-navigation, structure-modification, geometric-analysis
chemoinformatics/ — molecular-io, molecular-descriptors, similarity-searching, substructure-search, virtual-screening, admet-prediction, reaction-enumeration
ClawBio:
struct-predictor — Local AlphaFold/Boltz/Chai structure prediction with comparison
### Proteomics
bioSkills:
proteomics/ — data-import, peptide-identification, protein-inference, quantification, differential-abundance, dia-analysis, ptm-analysis, proteomics-qc, spectral-libraries
ClawBio:
proteomics-de — Proteomics differential expression
### Pathway Analysis & Gene Networks
bioSkills:
pathway-analysis/ — go-enrichment, gsea, kegg-pathways, reactome-pathways, wikipathways, enrichment-visualization
gene-regulatory-networks/ — scenic-regulons, coexpression-networks, differential-networks, multiomics-grn, perturbation-simulation
### Immunoinformatics
bioSkills:
immunoinformatics/ — mhc-binding-prediction, epitope-prediction, neoantigen-prediction, immunogenicity-scoring, tcr-epitope-binding
tcr-bcr-analysis/ — mixcr-analysis, scirpy-analysis, immcantation-analysis, repertoire-visualization, vdjtools-analysis
### CRISPR & Genome Engineering
bioSkills:
crispr-screens/ — mageck-analysis, jacks-analysis, hit-calling, screen-qc, library-design, crispresso-editing, base-editing-analysis, batch-correction
genome-engineering/ — grna-design, off-target-prediction, hdr-template-design, base-editing-design, prime-editing-design
### Workflow Management
bioSkills:
workflow-management/ — snakemake-workflows, nextflow-pipelines, cwl-workflows, wdl-workflows
ClawBio:
repro-enforcer — Export any analysis as reproducibility bundle (Conda env + Singularity + checksums)
galaxy-bridge — Access 8,000+ Galaxy tools from usegalaxy.org
### Specialized Domains
bioSkills:
alternative-splicing/ — splicing-quantification, differential-splicing, isoform-switching, sashimi-plots, single-cell-splicing, splicing-qc
ecological-genomics/ — edna-metabarcoding, landscape-genomics, conservation-genetics, biodiversity-metrics, community-ecology, species-delimitation
epidemiological-genomics/ — pathogen-typing, variant-surveillance, phylodynamics, transmission-inference, amr-surveillance
liquid-biopsy/ — cfdna-preprocessing, ctdna-mutation-detection, fragment-analysis, tumor-fraction-estimation, methylation-based-detection, longitudinal-monitoring
epitranscriptomics/ — m6a-peak-calling, m6a-differential, m6anet-analysis, merip-preprocessing, modification-visualization
metabolomics/ — xcms-preprocessing, metabolite-annotation, normalization-qc, statistical-analysis, pathway-mapping, lipidomics, targeted-analysis, msdial-preprocessing
flow-cytometry/ — fcs-handling, gating-analysis, compensation-transformation, clustering-phenotyping, differential-analysis, cytometry-qc, doublet-detection, bead-normalization
systems-biology/ — flux-balance-analysis, metabolic-reconstruction, gene-essentiality, context-specific-models, model-curation
rna-structure/ — secondary-structure-prediction, ncrna-search, structure-probing
### Data Visualization & Reporting
bioSkills:
data-visualization/ — ggplot2-fundamentals, heatmaps-clustering, volcano-customization, circos-plots, genome-browser-tracks, interactive-visualization, multipanel-figures, network-visualization, upset-plots, color-palettes, specialized-omics-plots, genome-tracks
reporting/ — rmarkdown-reports, quarto-reports, jupyter-reports, automated-qc-reports, figure-export
ClawBio:
profile-report — Analysis profile reporting
data-extractor — Extract numerical data from scientific figure images (via vision)
lit-synthesizer — PubMed/bioRxiv search, summarization, citation graphs
pubmed-summariser — Gene/disease PubMed search with structured briefing
### Database Access
bioSkills:
database-access/ — entrez-search, entrez-fetch, entrez-link, blast-searches, local-blast, sra-data, geo-data, uniprot-access, batch-downloads, interaction-databases, sequence-similarity
ClawBio:
ukb-navigator — Semantic search across 12,000+ UK Biobank fields
clinical-trial-finder — Clinical trial discovery
### Experimental Design
bioSkills:
experimental-design/ — power-analysis, sample-size, batch-design, multiple-testing
### Machine Learning for Omics
bioSkills:
machine-learning/ — omics-classifiers, biomarker-discovery, survival-analysis, model-validation, prediction-explanation, atlas-mapping
ClawBio:
claw-semantic-sim — Semantic similarity index for disease literature (PubMedBERT)
omics-target-evidence-mapper — Aggregate target-level evidence across omics sources
## Environment Setup
These skills assume a bioinformatics workstation. Common dependencies:
```bash
# Python
pip install biopython pysam cyvcf2 pybedtools pyBigWig scikit-allel anndata scanpy mygene
# R/Bioconductor
Rscript -e 'BiocManager::install(c("DESeq2","edgeR","Seurat","clusterProfiler","methylKit"))'
# CLI tools (Ubuntu/Debian)
sudo apt install samtools bcftools ncbi-blast+ minimap2 bedtools
# CLI tools (macOS)
brew install samtools bcftools blast minimap2 bedtools
# Or via Conda (recommended for reproducibility)
conda install -c bioconda samtools bcftools blast minimap2 bedtools fastp kraken2
```
## Pitfalls
- The fetched skills are NOT in Hermes SKILL.md format. They use their own structure (bioSkills: code pattern cookbooks; ClawBio: README + Python scripts). Read them as expert reference material.
- bioSkills are reference guides — they show correct parameters and code patterns but aren't executable pipelines.
- ClawBio skills are executable — many have `--demo` flags and can be run directly.
- Both repos assume bioinformatics tools are installed. Check prerequisites before running pipelines.
- For ClawBio, run `pip install -r requirements.txt` in the cloned repo first.
- Genomic data files can be very large. Be mindful of disk space when downloading reference genomes, SRA datasets, or building indices.

116
website/docs/user-guide/skills/optional/research/research-domain-intel.md Normal file
View file

@ -0,0 +1,116 @@
---
title: "Domain Intel — Passive domain reconnaissance using Python stdlib"
sidebar_label: "Domain Intel"
description: "Passive domain reconnaissance using Python stdlib"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Domain Intel
Passive domain reconnaissance using Python stdlib. Subdomain discovery, SSL certificate inspection, WHOIS lookups, DNS records, domain availability checks, and bulk multi-domain analysis. No API keys required.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/research/domain-intel` |
| Path | `optional-skills/research/domain-intel` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Domain Intelligence — Passive OSINT
Passive domain reconnaissance using only Python stdlib.
**Zero dependencies. Zero API keys. Works on Linux, macOS, and Windows.**
## Helper script
This skill includes `scripts/domain_intel.py` — a complete CLI tool for all domain intelligence operations.
```bash
# Subdomain discovery via Certificate Transparency logs
python3 SKILL_DIR/scripts/domain_intel.py subdomains example.com
# SSL certificate inspection (expiry, cipher, SANs, issuer)
python3 SKILL_DIR/scripts/domain_intel.py ssl example.com
# WHOIS lookup (registrar, dates, name servers — 100+ TLDs)
python3 SKILL_DIR/scripts/domain_intel.py whois example.com
# DNS records (A, AAAA, MX, NS, TXT, CNAME)
python3 SKILL_DIR/scripts/domain_intel.py dns example.com
# Domain availability check (passive: DNS + WHOIS + SSL signals)
python3 SKILL_DIR/scripts/domain_intel.py available coolstartup.io
# Bulk analysis — multiple domains, multiple checks in parallel
python3 SKILL_DIR/scripts/domain_intel.py bulk example.com github.com google.com
python3 SKILL_DIR/scripts/domain_intel.py bulk example.com github.com --checks ssl,dns
```
`SKILL_DIR` is the directory containing this SKILL.md file. All output is structured JSON.
## Available commands
| Command | What it does | Data source |
|---------|-------------|-------------|
| `subdomains` | Find subdomains from certificate logs | crt.sh (HTTPS) |
| `ssl` | Inspect TLS certificate details | Direct TCP:443 to target |
| `whois` | Registration info, registrar, dates | WHOIS servers (TCP:43) |
| `dns` | A, AAAA, MX, NS, TXT, CNAME records | System DNS + Google DoH |
| `available` | Check if domain is registered | DNS + WHOIS + SSL signals |
| `bulk` | Run multiple checks on multiple domains | All of the above |
## When to use this vs built-in tools
- **Use this skill** for infrastructure questions: subdomains, SSL certs, WHOIS, DNS records, availability
- **Use `web_search`** for general research about what a domain/company does
- **Use `web_extract`** to get the actual content of a webpage
- **Use `terminal` with `curl -I`** for a simple "is this URL reachable" check
| Task | Better tool | Why |
|------|-------------|-----|
| "What does example.com do?" | `web_extract` | Gets page content, not DNS/WHOIS data |
| "Find info about a company" | `web_search` | General research, not domain-specific |
| "Is this website safe?" | `web_search` | Reputation checks need web context |
| "Check if a URL is reachable" | `terminal` with `curl -I` | Simple HTTP check |
| "Find subdomains of X" | **This skill** | Only passive source for this |
| "When does the SSL cert expire?" | **This skill** | Built-in tools can't inspect TLS |
| "Who registered this domain?" | **This skill** | WHOIS data not in web search |
| "Is coolstartup.io available?" | **This skill** | Passive availability via DNS+WHOIS+SSL |
## Platform compatibility
Pure Python stdlib (`socket`, `ssl`, `urllib`, `json`, `concurrent.futures`).
Works identically on Linux, macOS, and Windows with no dependencies.
- **crt.sh queries** use HTTPS (port 443) — works behind most firewalls
- **WHOIS queries** use TCP port 43 — may be blocked on restrictive networks
- **DNS queries** use Google DoH (HTTPS) for MX/NS/TXT — firewall-friendly
- **SSL checks** connect to the target on port 443 — the only "active" operation
## Data sources
All queries are **passive** — no port scanning, no vulnerability testing:
- **crt.sh** — Certificate Transparency logs (subdomain discovery, HTTPS only)
- **WHOIS servers** — Direct TCP to 100+ authoritative TLD registrars
- **Google DNS-over-HTTPS** — MX, NS, TXT, CNAME resolution (firewall-friendly)
- **System DNS** — A/AAAA record resolution
- **SSL check** is the only "active" operation (TCP connection to target:443)
## Notes
- WHOIS queries use TCP port 43 — may be blocked on restrictive networks
- Some WHOIS servers redact registrant info (GDPR) — mention this to the user
- crt.sh can be slow for very popular domains (thousands of certs) — set reasonable expectations
- The availability check is heuristic-based (3 passive signals) — not authoritative like a registrar API
---
*Contributed by [@FurkanL0](https://github.com/FurkanL0)*

236
website/docs/user-guide/skills/optional/research/research-drug-discovery.md Normal file
View file

@ -0,0 +1,236 @@
---
title: "Drug Discovery — Pharmaceutical research assistant for drug discovery workflows"
sidebar_label: "Drug Discovery"
description: "Pharmaceutical research assistant for drug discovery workflows"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Drug Discovery
Pharmaceutical research assistant for drug discovery workflows. Search bioactive compounds on ChEMBL, calculate drug-likeness (Lipinski Ro5, QED, TPSA, synthetic accessibility), look up drug-drug interactions via OpenFDA, interpret ADMET profiles, and assist with lead optimization. Use for medicinal chemistry questions, molecule property analysis, clinical pharmacology, and open-science drug research.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/research/drug-discovery` |
| Path | `optional-skills/research/drug-discovery` |
| Version | `1.0.0` |
| Author | bennytimz |
| License | MIT |
| Tags | `science`, `chemistry`, `pharmacology`, `research`, `health` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Drug Discovery & Pharmaceutical Research
You are an expert pharmaceutical scientist and medicinal chemist with deep
knowledge of drug discovery, cheminformatics, and clinical pharmacology.
Use this skill for all pharma/chemistry research tasks.
## Core Workflows
### 1 — Bioactive Compound Search (ChEMBL)
Search ChEMBL (the world's largest open bioactivity database) for compounds
by target, activity, or molecule name. No API key required.
```bash
# Search compounds by target name (e.g. "EGFR", "COX-2", "ACE")
TARGET="$1"
ENCODED=$(python3 -c "import urllib.parse,sys; print(urllib.parse.quote(sys.argv[1]))" "$TARGET")
curl -s "https://www.ebi.ac.uk/chembl/api/data/target/search?q=${ENCODED}&format=json" \
| python3 -c "
import json,sys
data=json.load(sys.stdin)
targets=data.get('targets',[])[:5]
for t in targets:
print(f\"ChEMBL ID : {t.get('target_chembl_id')}\")
print(f\"Name : {t.get('pref_name')}\")
print(f\"Type : {t.get('target_type')}\")
print()
"
```
```bash
# Get bioactivity data for a ChEMBL target ID
TARGET_ID="$1" # e.g. CHEMBL203
curl -s "https://www.ebi.ac.uk/chembl/api/data/activity?target_chembl_id=${TARGET_ID}&pchembl_value__gte=6&limit=10&format=json" \
| python3 -c "
import json,sys
data=json.load(sys.stdin)
acts=data.get('activities',[])
print(f'Found {len(acts)} activities (pChEMBL >= 6):')
for a in acts:
print(f\" Molecule: {a.get('molecule_chembl_id')} | {a.get('standard_type')}: {a.get('standard_value')} {a.get('standard_units')} | pChEMBL: {a.get('pchembl_value')}\")
"
```
```bash
# Look up a specific molecule by ChEMBL ID
MOL_ID="$1" # e.g. CHEMBL25 (aspirin)
curl -s "https://www.ebi.ac.uk/chembl/api/data/molecule/${MOL_ID}?format=json" \
| python3 -c "
import json,sys
m=json.load(sys.stdin)
props=m.get('molecule_properties',{}) or {}
print(f\"Name : {m.get('pref_name','N/A')}\")
print(f\"SMILES : {m.get('molecule_structures',{}).get('canonical_smiles','N/A') if m.get('molecule_structures') else 'N/A'}\")
print(f\"MW : {props.get('full_mwt','N/A')} Da\")
print(f\"LogP : {props.get('alogp','N/A')}\")
print(f\"HBD : {props.get('hbd','N/A')}\")
print(f\"HBA : {props.get('hba','N/A')}\")
print(f\"TPSA : {props.get('psa','N/A')} Ų\")
print(f\"Ro5 violations: {props.get('num_ro5_violations','N/A')}\")
print(f\"QED : {props.get('qed_weighted','N/A')}\")
"
```
### 2 — Drug-Likeness Calculation (Lipinski Ro5 + Veber)
Assess any molecule against established oral bioavailability rules using
PubChem's free property API — no RDKit install needed.
```bash
COMPOUND="$1"
ENCODED=$(python3 -c "import urllib.parse,sys; print(urllib.parse.quote(sys.argv[1]))" "$COMPOUND")
curl -s "https://pubchem.ncbi.nlm.nih.gov/rest/pug/compound/name/${ENCODED}/property/MolecularWeight,XLogP,HBondDonorCount,HBondAcceptorCount,RotatableBondCount,TPSA,InChIKey/JSON" \
| python3 -c "
import json,sys
data=json.load(sys.stdin)
props=data['PropertyTable']['Properties'][0]
mw = float(props.get('MolecularWeight', 0))
logp = float(props.get('XLogP', 0))
hbd = int(props.get('HBondDonorCount', 0))
hba = int(props.get('HBondAcceptorCount', 0))
rot = int(props.get('RotatableBondCount', 0))
tpsa = float(props.get('TPSA', 0))
print('=== Lipinski Rule of Five (Ro5) ===')
print(f' MW {mw:.1f} Da {\"✓\" if mw<=500 else \"✗ VIOLATION (>500)\"}')
print(f' LogP {logp:.2f} {\"✓\" if logp<=5 else \"✗ VIOLATION (>5)\"}')
print(f' HBD {hbd} {\"✓\" if hbd<=5 else \"✗ VIOLATION (>5)\"}')
print(f' HBA {hba} {\"✓\" if hba<=10 else \"✗ VIOLATION (>10)\"}')
viol = sum([mw>500, logp>5, hbd>5, hba>10])
print(f' Violations: {viol}/4 {\"→ Likely orally bioavailable\" if viol<=1 else \"→ Poor oral bioavailability predicted\"}')
print()
print('=== Veber Oral Bioavailability Rules ===')
print(f' TPSA {tpsa:.1f} Ų {\"✓\" if tpsa<=140 else \"✗ VIOLATION (>140)\"}')
print(f' Rot. bonds {rot} {\"✓\" if rot<=10 else \"✗ VIOLATION (>10)\"}')
print(f' Both rules met: {\"Yes → good oral absorption predicted\" if tpsa<=140 and rot<=10 else \"No → reduced oral absorption\"}')
"
```
### 3 — Drug Interaction & Safety Lookup (OpenFDA)
```bash
DRUG="$1"
ENCODED=$(python3 -c "import urllib.parse,sys; print(urllib.parse.quote(sys.argv[1]))" "$DRUG")
curl -s "https://api.fda.gov/drug/label.json?search=drug_interactions:\"${ENCODED}\"&limit=3" \
| python3 -c "
import json,sys
data=json.load(sys.stdin)
results=data.get('results',[])
if not results:
print('No interaction data found in FDA labels.')
sys.exit()
for r in results[:2]:
brand=r.get('openfda',{}).get('brand_name',['Unknown'])[0]
generic=r.get('openfda',{}).get('generic_name',['Unknown'])[0]
interactions=r.get('drug_interactions',['N/A'])[0]
print(f'--- {brand} ({generic}) ---')
print(interactions[:800])
print()
"
```
```bash
DRUG="$1"
ENCODED=$(python3 -c "import urllib.parse,sys; print(urllib.parse.quote(sys.argv[1]))" "$DRUG")
curl -s "https://api.fda.gov/drug/event.json?search=patient.drug.medicinalproduct:\"${ENCODED}\"&count=patient.reaction.reactionmeddrapt.exact&limit=10" \
| python3 -c "
import json,sys
data=json.load(sys.stdin)
results=data.get('results',[])
if not results:
print('No adverse event data found.')
sys.exit()
print(f'Top adverse events reported:')
for r in results[:10]:
print(f\" {r['count']:>5}x {r['term']}\")
"
```
### 4 — PubChem Compound Search
```bash
COMPOUND="$1"
ENCODED=$(python3 -c "import urllib.parse,sys; print(urllib.parse.quote(sys.argv[1]))" "$COMPOUND")
CID=$(curl -s "https://pubchem.ncbi.nlm.nih.gov/rest/pug/compound/name/${ENCODED}/cids/TXT" | head -1 | tr -d '[:space:]')
echo "PubChem CID: $CID"
curl -s "https://pubchem.ncbi.nlm.nih.gov/rest/pug/compound/cid/${CID}/property/IsomericSMILES,InChIKey,IUPACName/JSON" \
| python3 -c "
import json,sys
p=json.load(sys.stdin)['PropertyTable']['Properties'][0]
print(f\"IUPAC Name : {p.get('IUPACName','N/A')}\")
print(f\"SMILES : {p.get('IsomericSMILES','N/A')}\")
print(f\"InChIKey : {p.get('InChIKey','N/A')}\")
"
```
### 5 — Target & Disease Literature (OpenTargets)
```bash
GENE="$1"
curl -s -X POST "https://api.platform.opentargets.org/api/v4/graphql" \
-H "Content-Type: application/json" \
-d "{\"query\":\"{ search(queryString: \\\"${GENE}\\\", entityNames: [\\\"target\\\"], page: {index: 0, size: 1}) { hits { id score object { ... on Target { id approvedSymbol approvedName associatedDiseases(page: {index: 0, size: 5}) { count rows { score disease { id name } } } } } } } }\"}" \
| python3 -c "
import json,sys
data=json.load(sys.stdin)
hits=data.get('data',{}).get('search',{}).get('hits',[])
if not hits:
print('Target not found.')
sys.exit()
obj=hits[0]['object']
print(f\"Target: {obj.get('approvedSymbol')} — {obj.get('approvedName')}\")
assoc=obj.get('associatedDiseases',{})
print(f\"Associated with {assoc.get('count',0)} diseases. Top associations:\")
for row in assoc.get('rows',[]):
print(f\" Score {row['score']:.3f} | {row['disease']['name']}\")
"
```
## Reasoning Guidelines
When analysing drug-likeness or molecular properties, always:
1. **State raw values first** — MW, LogP, HBD, HBA, TPSA, RotBonds
2. **Apply rule sets** — Ro5 (Lipinski), Veber, Ghose filter where relevant
3. **Flag liabilities** — metabolic hotspots, hERG risk, high TPSA for CNS penetration
4. **Suggest optimizations** — bioisosteric replacements, prodrug strategies, ring truncation
5. **Cite the source API** — ChEMBL, PubChem, OpenFDA, or OpenTargets
For ADMET questions, reason through Absorption, Distribution, Metabolism, Excretion, Toxicity systematically. See references/ADMET_REFERENCE.md for detailed guidance.
## Important Notes
- All APIs are free, public, require no authentication
- ChEMBL rate limits: add sleep 1 between batch requests
- FDA data reflects reported adverse events, not necessarily causation
- Always recommend consulting a licensed pharmacist or physician for clinical decisions
## Quick Reference
| Task | API | Endpoint |
|------|-----|----------|
| Find target | ChEMBL | `/api/data/target/search?q=` |
| Get bioactivity | ChEMBL | `/api/data/activity?target_chembl_id=` |
| Molecule properties | PubChem | `/rest/pug/compound/name/{name}/property/` |
| Drug interactions | OpenFDA | `/drug/label.json?search=drug_interactions:` |
| Adverse events | OpenFDA | `/drug/event.json?search=...&count=reaction` |
| Gene-disease | OpenTargets | GraphQL POST `/api/v4/graphql` |

254
website/docs/user-guide/skills/optional/research/research-duckduckgo-search.md Normal file
View file

@ -0,0 +1,254 @@
---
title: "Duckduckgo Search — Free web search via DuckDuckGo — text, news, images, videos"
sidebar_label: "Duckduckgo Search"
description: "Free web search via DuckDuckGo — text, news, images, videos"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Duckduckgo Search
Free web search via DuckDuckGo — text, news, images, videos. No API key needed. Prefer the `ddgs` CLI when installed; use the Python DDGS library only after verifying that `ddgs` is available in the current runtime.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/research/duckduckgo-search` |
| Path | `optional-skills/research/duckduckgo-search` |
| Version | `1.3.0` |
| Author | gamedevCloudy |
| License | MIT |
| Tags | `search`, `duckduckgo`, `web-search`, `free`, `fallback` |
| Related skills | [`arxiv`](/docs/user-guide/skills/bundled/research/research-arxiv) |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# DuckDuckGo Search
Free web search using DuckDuckGo. **No API key required.**
Preferred when `web_search` is unavailable or unsuitable (for example when `FIRECRAWL_API_KEY` is not set). Can also be used as a standalone search path when DuckDuckGo results are specifically desired.
## Detection Flow
Check what is actually available before choosing an approach:
```bash
# Check CLI availability
command -v ddgs >/dev/null && echo "DDGS_CLI=installed" || echo "DDGS_CLI=missing"
```
Decision tree:
1. If `ddgs` CLI is installed, prefer `terminal` + `ddgs`
2. If `ddgs` CLI is missing, do not assume `execute_code` can import `ddgs`
3. If the user wants DuckDuckGo specifically, install `ddgs` first in the relevant environment
4. Otherwise fall back to built-in web/browser tools
Important runtime note:
- Terminal and `execute_code` are separate runtimes
- A successful shell install does not guarantee `execute_code` can import `ddgs`
- Never assume third-party Python packages are preinstalled inside `execute_code`
## Installation
Install `ddgs` only when DuckDuckGo search is specifically needed and the runtime does not already provide it.
```bash
# Python package + CLI entrypoint
pip install ddgs
# Verify CLI
ddgs --help
```
If a workflow depends on Python imports, verify that same runtime can import `ddgs` before using `from ddgs import DDGS`.
## Method 1: CLI Search (Preferred)
Use the `ddgs` command via `terminal` when it exists. This is the preferred path because it avoids assuming the `execute_code` sandbox has the `ddgs` Python package installed.
```bash
# Text search
ddgs text -q "python async programming" -m 5
# News search
ddgs news -q "artificial intelligence" -m 5
# Image search
ddgs images -q "landscape photography" -m 10
# Video search
ddgs videos -q "python tutorial" -m 5
# With region filter
ddgs text -q "best restaurants" -m 5 -r us-en
# Recent results only (d=day, w=week, m=month, y=year)
ddgs text -q "latest AI news" -m 5 -t w
# JSON output for parsing
ddgs text -q "fastapi tutorial" -m 5 -o json
```
### CLI Flags
| Flag | Description | Example |
|------|-------------|---------|
| `-q` | Query — **required** | `-q "search terms"` |
| `-m` | Max results | `-m 5` |
| `-r` | Region | `-r us-en` |
| `-t` | Time limit | `-t w` (week) |
| `-s` | Safe search | `-s off` |
| `-o` | Output format | `-o json` |
## Method 2: Python API (Only After Verification)
Use the `DDGS` class in `execute_code` or another Python runtime only after verifying that `ddgs` is installed there. Do not assume `execute_code` includes third-party packages by default.
Safe wording:
- "Use `execute_code` with `ddgs` after installing or verifying the package if needed"
Avoid saying:
- "`execute_code` includes `ddgs`"
- "DuckDuckGo search works by default in `execute_code`"
**Important:** `max_results` must always be passed as a **keyword argument** — positional usage raises an error on all methods.
### Text Search
Best for: general research, companies, documentation.
```python
from ddgs import DDGS
with DDGS() as ddgs:
for r in ddgs.text("python async programming", max_results=5):
print(r["title"])
print(r["href"])
print(r.get("body", "")[:200])
print()
```
Returns: `title`, `href`, `body`
### News Search
Best for: current events, breaking news, latest updates.
```python
from ddgs import DDGS
with DDGS() as ddgs:
for r in ddgs.news("AI regulation 2026", max_results=5):
print(r["date"], "-", r["title"])
print(r.get("source", ""), "|", r["url"])
print(r.get("body", "")[:200])
print()
```
Returns: `date`, `title`, `body`, `url`, `image`, `source`
### Image Search
Best for: visual references, product images, diagrams.
```python
from ddgs import DDGS
with DDGS() as ddgs:
for r in ddgs.images("semiconductor chip", max_results=5):
print(r["title"])
print(r["image"])
print(r.get("thumbnail", ""))
print(r.get("source", ""))
print()
```
Returns: `title`, `image`, `thumbnail`, `url`, `height`, `width`, `source`
### Video Search
Best for: tutorials, demos, explainers.
```python
from ddgs import DDGS
with DDGS() as ddgs:
for r in ddgs.videos("FastAPI tutorial", max_results=5):
print(r["title"])
print(r.get("content", ""))
print(r.get("duration", ""))
print(r.get("provider", ""))
print(r.get("published", ""))
print()
```
Returns: `title`, `content`, `description`, `duration`, `provider`, `published`, `statistics`, `uploader`
### Quick Reference
| Method | Use When | Key Fields |
|--------|----------|------------|
| `text()` | General research, companies | title, href, body |
| `news()` | Current events, updates | date, title, source, body, url |
| `images()` | Visuals, diagrams | title, image, thumbnail, url |
| `videos()` | Tutorials, demos | title, content, duration, provider |
## Workflow: Search then Extract
DuckDuckGo returns titles, URLs, and snippets — not full page content. To get full page content, search first and then extract the most relevant URL with `web_extract`, browser tools, or curl.
CLI example:
```bash
ddgs text -q "fastapi deployment guide" -m 3 -o json
```
Python example, only after verifying `ddgs` is installed in that runtime:
```python
from ddgs import DDGS
with DDGS() as ddgs:
results = list(ddgs.text("fastapi deployment guide", max_results=3))
for r in results:
print(r["title"], "->", r["href"])
```
Then extract the best URL with `web_extract` or another content-retrieval tool.
## Limitations
- **Rate limiting**: DuckDuckGo may throttle after many rapid requests. Add a short delay between searches if needed.
- **No content extraction**: `ddgs` returns snippets, not full page content. Use `web_extract`, browser tools, or curl for the full article/page.
- **Results quality**: Generally good but less configurable than Firecrawl's search.
- **Availability**: DuckDuckGo may block requests from some cloud IPs. If searches return empty, try different keywords or wait a few seconds.
- **Field variability**: Return fields may vary between results or `ddgs` versions. Use `.get()` for optional fields to avoid `KeyError`.
- **Separate runtimes**: A successful `ddgs` install in terminal does not automatically mean `execute_code` can import it.
## Troubleshooting
| Problem | Likely Cause | What To Do |
|---------|--------------|------------|
| `ddgs: command not found` | CLI not installed in the shell environment | Install `ddgs`, or use built-in web/browser tools instead |
| `ModuleNotFoundError: No module named 'ddgs'` | Python runtime does not have the package installed | Do not use Python DDGS there until that runtime is prepared |
| Search returns nothing | Temporary rate limiting or poor query | Wait a few seconds, retry, or adjust the query |
| CLI works but `execute_code` import fails | Terminal and `execute_code` are different runtimes | Keep using CLI, or separately prepare the Python runtime |
## Pitfalls
- **`max_results` is keyword-only**: `ddgs.text("query", 5)` raises an error. Use `ddgs.text("query", max_results=5)`.
- **Do not assume the CLI exists**: Check `command -v ddgs` before using it.
- **Do not assume `execute_code` can import `ddgs`**: `from ddgs import DDGS` may fail with `ModuleNotFoundError` unless that runtime was prepared separately.
- **Package name**: The package is `ddgs` (previously `duckduckgo-search`). Install with `pip install ddgs`.
- **Don't confuse `-q` and `-m`** (CLI): `-q` is for the query, `-m` is for max results count.
- **Empty results**: If `ddgs` returns nothing, it may be rate-limited. Wait a few seconds and retry.
## Validated With
Validated examples against `ddgs==9.11.2` semantics. Skill guidance now treats CLI availability and Python import availability as separate concerns so the documented workflow matches actual runtime behavior.

231
website/docs/user-guide/skills/optional/research/research-gitnexus-explorer.md Normal file
View file

@ -0,0 +1,231 @@
---
title: "Gitnexus Explorer"
sidebar_label: "Gitnexus Explorer"
description: "Index a codebase with GitNexus and serve an interactive knowledge graph via web UI + Cloudflare tunnel"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Gitnexus Explorer
Index a codebase with GitNexus and serve an interactive knowledge graph via web UI + Cloudflare tunnel.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/research/gitnexus-explorer` |
| Path | `optional-skills/research/gitnexus-explorer` |
| Version | `1.0.0` |
| Author | Hermes Agent + Teknium |
| License | MIT |
| Tags | `gitnexus`, `code-intelligence`, `knowledge-graph`, `visualization` |
| Related skills | [`native-mcp`](/docs/user-guide/skills/bundled/mcp/mcp-native-mcp), [`codebase-inspection`](/docs/user-guide/skills/bundled/github/github-codebase-inspection) |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# GitNexus Explorer
Index any codebase into a knowledge graph and serve an interactive web UI for exploring
symbols, call chains, clusters, and execution flows. Tunneled via Cloudflare for remote access.
## When to Use
- User wants to visually explore a codebase's architecture
- User asks for a knowledge graph / dependency graph of a repo
- User wants to share an interactive codebase explorer with someone
## Prerequisites
- **Node.js** (v18+) — required for GitNexus and the proxy
- **git** — repo must have a `.git` directory
- **cloudflared** — for tunneling (auto-installed to ~/.local/bin if missing)
## Size Warning
The web UI renders all nodes in the browser. Repos under ~5,000 files work well. Large
repos (30k+ nodes) will be sluggish or crash the browser tab. The CLI/MCP tools work
at any scale — only the web visualization has this limit.
## Steps
### 1. Clone and Build GitNexus (one-time setup)
```bash
GITNEXUS_DIR="${GITNEXUS_DIR:-$HOME/.local/share/gitnexus}"
if [ ! -d "$GITNEXUS_DIR/gitnexus-web/dist" ]; then
git clone https://github.com/abhigyanpatwari/GitNexus.git "$GITNEXUS_DIR"
cd "$GITNEXUS_DIR/gitnexus-shared" && npm install && npm run build
cd "$GITNEXUS_DIR/gitnexus-web" && npm install
fi
```
### 2. Patch the Web UI for Remote Access
The web UI defaults to `localhost:4747` for API calls. Patch it to use same-origin
so it works through a tunnel/proxy:
**File: `$GITNEXUS_DIR/gitnexus-web/src/config/ui-constants.ts`**
Change:
```typescript
export const DEFAULT_BACKEND_URL = 'http://localhost:4747';
```
To:
```typescript
export const DEFAULT_BACKEND_URL = typeof window !== 'undefined' && window.location.hostname !== 'localhost' ? window.location.origin : 'http://localhost:4747';
```
**File: `$GITNEXUS_DIR/gitnexus-web/vite.config.ts`**
Add `allowedHosts: true` inside the `server: { }` block (only needed if running dev
mode instead of production build):
```typescript
server: {
allowedHosts: true,
// ... existing config
},
```
Then build the production bundle:
```bash
cd "$GITNEXUS_DIR/gitnexus-web" && npx vite build
```
### 3. Index the Target Repo
```bash
cd /path/to/target-repo
npx gitnexus analyze --skip-agents-md
rm -rf .claude/ # remove Claude Code-specific artifacts
```
Add `--embeddings` for semantic search (slower — minutes instead of seconds).
The index lives in `.gitnexus/` inside the repo (auto-gitignored).
### 4. Create the Proxy Script
Write this to a file (e.g., `$GITNEXUS_DIR/proxy.mjs`). It serves the production
web UI and proxies `/api/*` to the GitNexus backend — same origin, no CORS issues,
no sudo, no nginx.
```javascript
import http from 'node:http';
import fs from 'node:fs';
import path from 'node:path';
const API_PORT = parseInt(process.env.API_PORT || '4747');
const DIST_DIR = process.argv[2] || './dist';
const PORT = parseInt(process.argv[3] || '8888');
const MIME = {
'.html': 'text/html', '.js': 'application/javascript', '.css': 'text/css',
'.json': 'application/json', '.png': 'image/png', '.svg': 'image/svg+xml',
'.ico': 'image/x-icon', '.woff2': 'font/woff2', '.woff': 'font/woff',
'.wasm': 'application/wasm',
};
function proxyToApi(req, res) {
const opts = {
hostname: '127.0.0.1', port: API_PORT,
path: req.url, method: req.method, headers: req.headers,
};
const proxy = http.request(opts, (upstream) => {
res.writeHead(upstream.statusCode, upstream.headers);
upstream.pipe(res, { end: true });
});
proxy.on('error', () => { res.writeHead(502); res.end('Backend unavailable'); });
req.pipe(proxy, { end: true });
}
function serveStatic(req, res) {
let filePath = path.join(DIST_DIR, req.url === '/' ? 'index.html' : req.url.split('?')[0]);
if (!fs.existsSync(filePath)) filePath = path.join(DIST_DIR, 'index.html');
const ext = path.extname(filePath);
const mime = MIME[ext] || 'application/octet-stream';
try {
const data = fs.readFileSync(filePath);
res.writeHead(200, { 'Content-Type': mime, 'Cache-Control': 'public, max-age=3600' });
res.end(data);
} catch { res.writeHead(404); res.end('Not found'); }
}
http.createServer((req, res) => {
if (req.url.startsWith('/api')) proxyToApi(req, res);
else serveStatic(req, res);
}).listen(PORT, () => console.log(`GitNexus proxy on http://localhost:${PORT}`));
```
### 5. Start the Services
```bash
# Terminal 1: GitNexus backend API
npx gitnexus serve &
# Terminal 2: Proxy (web UI + API on one port)
node "$GITNEXUS_DIR/proxy.mjs" "$GITNEXUS_DIR/gitnexus-web/dist" 8888 &
```
Verify: `curl -s http://localhost:8888/api/repos` should return the indexed repo(s).
### 6. Tunnel with Cloudflare (optional — for remote access)
```bash
# Install cloudflared if needed (no sudo)
if ! command -v cloudflared &>/dev/null; then
mkdir -p ~/.local/bin
curl -sL https://github.com/cloudflare/cloudflared/releases/latest/download/cloudflared-linux-amd64 \
-o ~/.local/bin/cloudflared
chmod +x ~/.local/bin/cloudflared
export PATH="$HOME/.local/bin:$PATH"
fi
# Start tunnel (--config /dev/null avoids conflicts with existing named tunnels)
cloudflared tunnel --config /dev/null --url http://localhost:8888 --no-autoupdate --protocol http2
```
The tunnel URL (e.g., `https://random-words.trycloudflare.com`) is printed to stderr.
Share it — anyone with the link can explore the graph.
### 7. Cleanup
```bash
# Stop services
pkill -f "gitnexus serve"
pkill -f "proxy.mjs"
pkill -f cloudflared
# Remove index from the target repo
cd /path/to/target-repo
npx gitnexus clean
rm -rf .claude/
```
## Pitfalls
- **`--config /dev/null` is required for cloudflared** if the user has an existing
named tunnel config at `~/.cloudflared/config.yml`. Without it, the catch-all
ingress rule in the config returns 404 for all quick tunnel requests.
- **Production build is mandatory for tunneling.** The Vite dev server blocks
non-localhost hosts by default (`allowedHosts`). The production build + Node
proxy avoids this entirely.
- **The web UI does NOT create `.claude/` or `CLAUDE.md`.** Those are created by
`npx gitnexus analyze`. Use `--skip-agents-md` to suppress the markdown files,
then `rm -rf .claude/` for the rest. These are Claude Code integrations that
hermes-agent users don't need.
- **Browser memory limit.** The web UI loads the entire graph into browser memory.
Repos with 5k+ files may be sluggish. 30k+ files will likely crash the tab.
- **Embeddings are optional.** `--embeddings` enables semantic search but takes
minutes on large repos. Skip it for quick exploration; add it if you want
natural language queries via the AI chat panel.
- **Multiple repos.** `gitnexus serve` serves ALL indexed repos. Index several
repos, start serve once, and the web UI lets you switch between them.

408
website/docs/user-guide/skills/optional/research/research-parallel-cli.md Normal file
View file

@ -0,0 +1,408 @@
---
title: "Parallel Cli"
sidebar_label: "Parallel Cli"
description: "Optional vendor skill for Parallel CLI — agent-native web search, extraction, deep research, enrichment, FindAll, and monitoring"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Parallel Cli
Optional vendor skill for Parallel CLI — agent-native web search, extraction, deep research, enrichment, FindAll, and monitoring. Prefer JSON output and non-interactive flows.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/research/parallel-cli` |
| Path | `optional-skills/research/parallel-cli` |
| Version | `1.1.0` |
| Author | Hermes Agent |
| License | MIT |
| Tags | `Research`, `Web`, `Search`, `Deep-Research`, `Enrichment`, `CLI` |
| Related skills | [`duckduckgo-search`](/docs/user-guide/skills/optional/research/research-duckduckgo-search), [`mcporter`](/docs/user-guide/skills/optional/mcp/mcp-mcporter) |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Parallel CLI
Use `parallel-cli` when the user explicitly wants Parallel, or when a terminal-native workflow would benefit from Parallel's vendor-specific stack for web search, extraction, deep research, enrichment, entity discovery, or monitoring.
This is an optional third-party workflow, not a Hermes core capability.
Important expectations:
- Parallel is a paid service with a free tier, not a fully free local tool.
- It overlaps with Hermes native `web_search` / `web_extract`, so do not prefer it by default for ordinary lookups.
- Prefer this skill when the user mentions Parallel specifically or needs capabilities like Parallel's enrichment, FindAll, or monitor workflows.
`parallel-cli` is designed for agents:
- JSON output via `--json`
- Non-interactive command execution
- Async long-running jobs with `--no-wait`, `status`, and `poll`
- Context chaining with `--previous-interaction-id`
- Search, extract, research, enrichment, entity discovery, and monitoring in one CLI
## When to use it
Prefer this skill when:
- The user explicitly mentions Parallel or `parallel-cli`
- The task needs richer workflows than a simple one-shot search/extract pass
- You need async deep research jobs that can be launched and polled later
- You need structured enrichment, FindAll entity discovery, or monitoring
Prefer Hermes native `web_search` / `web_extract` for quick one-off lookups when Parallel is not specifically requested.
## Installation
Try the least invasive install path available for the environment.
### Homebrew
```bash
brew install parallel-web/tap/parallel-cli
```
### npm
```bash
npm install -g parallel-web-cli
```
### Python package
```bash
pip install "parallel-web-tools[cli]"
```
### Standalone installer
```bash
curl -fsSL https://parallel.ai/install.sh | bash
```
If you want an isolated Python install, `pipx` can also work:
```bash
pipx install "parallel-web-tools[cli]"
pipx ensurepath
```
## Authentication
Interactive login:
```bash
parallel-cli login
```
Headless / SSH / CI:
```bash
parallel-cli login --device
```
API key environment variable:
```bash
export PARALLEL_API_KEY="***"
```
Verify current auth status:
```bash
parallel-cli auth
```
If auth requires browser interaction, run with `pty=true`.
## Core rule set
1. Always prefer `--json` when you need machine-readable output.
2. Prefer explicit arguments and non-interactive flows.
3. For long-running jobs, use `--no-wait` and then `status` / `poll`.
4. Cite only URLs returned by the CLI output.
5. Save large JSON outputs to a temp file when follow-up questions are likely.
6. Use background processes only for genuinely long-running workflows; otherwise run in foreground.
7. Prefer Hermes native tools unless the user wants Parallel specifically or needs Parallel-only workflows.
## Quick reference
```text
parallel-cli
├── auth
├── login
├── logout
├── search
├── extract / fetch
├── research run|status|poll|processors
├── enrich run|status|poll|plan|suggest|deploy
├── findall run|ingest|status|poll|result|enrich|extend|schema|cancel
└── monitor create|list|get|update|delete|events|event-group|simulate
```
## Common flags and patterns
Commonly useful flags:
- `--json` for structured output
- `--no-wait` for async jobs
- `--previous-interaction-id <id>` for follow-up tasks that reuse earlier context
- `--max-results <n>` for search result count
- `--mode one-shot|agentic` for search behavior
- `--include-domains domain1.com,domain2.com`
- `--exclude-domains domain1.com,domain2.com`
- `--after-date YYYY-MM-DD`
Read from stdin when convenient:
```bash
echo "What is the latest funding for Anthropic?" | parallel-cli search - --json
echo "Research question" | parallel-cli research run - --json
```
## Search
Use for current web lookups with structured results.
```bash
parallel-cli search "What is Anthropic's latest AI model?" --json
parallel-cli search "SEC filings for Apple" --include-domains sec.gov --json
parallel-cli search "bitcoin price" --after-date 2026-01-01 --max-results 10 --json
parallel-cli search "latest browser benchmarks" --mode one-shot --json
parallel-cli search "AI coding agent enterprise reviews" --mode agentic --json
```
Useful constraints:
- `--include-domains` to narrow trusted sources
- `--exclude-domains` to strip noisy domains
- `--after-date` for recency filtering
- `--max-results` when you need broader coverage
If you expect follow-up questions, save output:
```bash
parallel-cli search "latest React 19 changes" --json -o /tmp/react-19-search.json
```
When summarizing results:
- lead with the answer
- include dates, names, and concrete facts
- cite only returned sources
- avoid inventing URLs or source titles
## Extraction
Use to pull clean content or markdown from a URL.
```bash
parallel-cli extract https://example.com --json
parallel-cli extract https://company.com --objective "Find pricing info" --json
parallel-cli extract https://example.com --full-content --json
parallel-cli fetch https://example.com --json
```
Use `--objective` when the page is broad and you only need one slice of information.
## Deep research
Use for deeper multi-step research tasks that may take time.
Common processor tiers:
- `lite` / `base` for faster, cheaper passes
- `core` / `pro` for more thorough synthesis
- `ultra` for the heaviest research jobs
### Synchronous
```bash
parallel-cli research run \
"Compare the leading AI coding agents by pricing, model support, and enterprise controls" \
--processor core \
--json
```
### Async launch + poll
```bash
parallel-cli research run \
"Compare the leading AI coding agents by pricing, model support, and enterprise controls" \
--processor ultra \
--no-wait \
--json
parallel-cli research status trun_xxx --json
parallel-cli research poll trun_xxx --json
parallel-cli research processors --json
```
### Context chaining / follow-up
```bash
parallel-cli research run "What are the top AI coding agents?" --json
parallel-cli research run \
"What enterprise controls does the top-ranked one offer?" \
--previous-interaction-id trun_xxx \
--json
```
Recommended Hermes workflow:
1. launch with `--no-wait --json`
2. capture the returned run/task ID
3. if the user wants to continue other work, keep moving
4. later call `status` or `poll`
5. summarize the final report with citations from the returned sources
## Enrichment
Use when the user has CSV/JSON/tabular inputs and wants additional columns inferred from web research.
### Suggest columns
```bash
parallel-cli enrich suggest "Find the CEO and annual revenue" --json
```
### Plan a config
```bash
parallel-cli enrich plan -o config.yaml
```
### Inline data
```bash
parallel-cli enrich run \
--data '[{"company": "Anthropic"}, {"company": "Mistral"}]' \
--intent "Find headquarters and employee count" \
--json
```
### Non-interactive file run
```bash
parallel-cli enrich run \
--source-type csv \
--source companies.csv \
--target enriched.csv \
--source-columns '[{"name": "company", "description": "Company name"}]' \
--intent "Find the CEO and annual revenue"
```
### YAML config run
```bash
parallel-cli enrich run config.yaml
```
### Status / polling
```bash
parallel-cli enrich status <task_group_id> --json
parallel-cli enrich poll <task_group_id> --json
```
Use explicit JSON arrays for column definitions when operating non-interactively.
Validate the output file before reporting success.
## FindAll
Use for web-scale entity discovery when the user wants a discovered dataset rather than a short answer.
```bash
parallel-cli findall run "Find AI coding agent startups with enterprise offerings" --json
parallel-cli findall run "AI startups in healthcare" -n 25 --json
parallel-cli findall status <run_id> --json
parallel-cli findall poll <run_id> --json
parallel-cli findall result <run_id> --json
parallel-cli findall schema <run_id> --json
```
This is a better fit than ordinary search when the user wants a discovered set of entities that can be reviewed, filtered, or enriched later.
## Monitor
Use for ongoing change detection over time.
```bash
parallel-cli monitor list --json
parallel-cli monitor get <monitor_id> --json
parallel-cli monitor events <monitor_id> --json
parallel-cli monitor delete <monitor_id> --json
```
Creation is usually the sensitive part because cadence and delivery matter:
```bash
parallel-cli monitor create --help
```
Use this when the user wants recurring tracking of a page or source rather than a one-time fetch.
## Recommended Hermes usage patterns
### Fast answer with citations
1. Run `parallel-cli search ... --json`
2. Parse titles, URLs, dates, excerpts
3. Summarize with inline citations from the returned URLs only
### URL investigation
1. Run `parallel-cli extract URL --json`
2. If needed, rerun with `--objective` or `--full-content`
3. Quote or summarize the extracted markdown
### Long research workflow
1. Run `parallel-cli research run ... --no-wait --json`
2. Store the returned ID
3. Continue other work or periodically poll
4. Summarize the final report with citations
### Structured enrichment workflow
1. Inspect the input file and columns
2. Use `enrich suggest` or provide explicit enriched columns
3. Run `enrich run`
4. Poll for completion if needed
5. Validate the output file before reporting success
## Error handling and exit codes
The CLI documents these exit codes:
- `0` success
- `2` bad input
- `3` auth error
- `4` API error
- `5` timeout
If you hit auth errors:
1. check `parallel-cli auth`
2. confirm `PARALLEL_API_KEY` or run `parallel-cli login` / `parallel-cli login --device`
3. verify `parallel-cli` is on `PATH`
## Maintenance
Check current auth / install state:
```bash
parallel-cli auth
parallel-cli --help
```
Update commands:
```bash
parallel-cli update
pip install --upgrade parallel-web-tools
parallel-cli config auto-update-check off
```
## Pitfalls
- Do not omit `--json` unless the user explicitly wants human-formatted output.
- Do not cite sources not present in the CLI output.
- `login` may require PTY/browser interaction.
- Prefer foreground execution for short tasks; do not overuse background processes.
- For large result sets, save JSON to `/tmp/*.json` instead of stuffing everything into context.
- Do not silently choose Parallel when Hermes native tools are already sufficient.
- Remember this is a vendor workflow that usually requires account auth and paid usage beyond the free tier.

459
website/docs/user-guide/skills/optional/research/research-qmd.md Normal file
View file

@ -0,0 +1,459 @@
---
title: "Qmd"
sidebar_label: "Qmd"
description: "Search personal knowledge bases, notes, docs, and meeting transcripts locally using qmd — a hybrid retrieval engine with BM25, vector search, and LLM reranking"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Qmd
Search personal knowledge bases, notes, docs, and meeting transcripts locally using qmd — a hybrid retrieval engine with BM25, vector search, and LLM reranking. Supports CLI and MCP integration.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/research/qmd` |
| Path | `optional-skills/research/qmd` |
| Version | `1.0.0` |
| Author | Hermes Agent + Teknium |
| License | MIT |
| Platforms | macos, linux |
| Tags | `Search`, `Knowledge-Base`, `RAG`, `Notes`, `MCP`, `Local-AI` |
| Related skills | [`obsidian`](/docs/user-guide/skills/bundled/note-taking/note-taking-obsidian), [`native-mcp`](/docs/user-guide/skills/bundled/mcp/mcp-native-mcp), [`arxiv`](/docs/user-guide/skills/bundled/research/research-arxiv) |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# QMD — Query Markup Documents
Local, on-device search engine for personal knowledge bases. Indexes markdown
notes, meeting transcripts, documentation, and any text-based files, then
provides hybrid search combining keyword matching, semantic understanding, and
LLM-powered reranking — all running locally with no cloud dependencies.
Created by [Tobi Lütke](https://github.com/tobi/qmd). MIT licensed.
## When to Use
- User asks to search their notes, docs, knowledge base, or meeting transcripts
- User wants to find something across a large collection of markdown/text files
- User wants semantic search ("find notes about X concept") not just keyword grep
- User has already set up qmd collections and wants to query them
- User asks to set up a local knowledge base or document search system
- Keywords: "search my notes", "find in my docs", "knowledge base", "qmd"
## Prerequisites
### Node.js >= 22 (required)
```bash
# Check version
node --version # must be >= 22
# macOS — install or upgrade via Homebrew
brew install node@22
# Linux — use NodeSource or nvm
curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash -
sudo apt-get install -y nodejs
# or with nvm:
nvm install 22 && nvm use 22
```
### SQLite with Extension Support (macOS only)
macOS system SQLite lacks extension loading. Install via Homebrew:
```bash
brew install sqlite
```
### Install qmd
```bash
npm install -g @tobilu/qmd
# or with Bun:
bun install -g @tobilu/qmd
```
First run auto-downloads 3 local GGUF models (~2GB total):
| Model | Purpose | Size |
|-------|---------|------|
| embeddinggemma-300M-Q8_0 | Vector embeddings | ~300MB |
| qwen3-reranker-0.6b-q8_0 | Result reranking | ~640MB |
| qmd-query-expansion-1.7B | Query expansion | ~1.1GB |
### Verify Installation
```bash
qmd --version
qmd status
```
## Quick Reference
| Command | What It Does | Speed |
|---------|-------------|-------|
| `qmd search "query"` | BM25 keyword search (no models) | ~0.2s |
| `qmd vsearch "query"` | Semantic vector search (1 model) | ~3s |
| `qmd query "query"` | Hybrid + reranking (all 3 models) | ~2-3s warm, ~19s cold |
| `qmd get <docid>` | Retrieve full document content | instant |
| `qmd multi-get "glob"` | Retrieve multiple files | instant |
| `qmd collection add <path> --name <n>` | Add a directory as a collection | instant |
| `qmd context add <path> "description"` | Add context metadata to improve retrieval | instant |
| `qmd embed` | Generate/update vector embeddings | varies |
| `qmd status` | Show index health and collection info | instant |
| `qmd mcp` | Start MCP server (stdio) | persistent |
| `qmd mcp --http --daemon` | Start MCP server (HTTP, warm models) | persistent |
## Setup Workflow
### 1. Add Collections
Point qmd at directories containing your documents:
```bash
# Add a notes directory
qmd collection add ~/notes --name notes
# Add project docs
qmd collection add ~/projects/myproject/docs --name project-docs
# Add meeting transcripts
qmd collection add ~/meetings --name meetings
# List all collections
qmd collection list
```
### 2. Add Context Descriptions
Context metadata helps the search engine understand what each collection
contains. This significantly improves retrieval quality:
```bash
qmd context add qmd://notes "Personal notes, ideas, and journal entries"
qmd context add qmd://project-docs "Technical documentation for the main project"
qmd context add qmd://meetings "Meeting transcripts and action items from team syncs"
```
### 3. Generate Embeddings
```bash
qmd embed
```
This processes all documents in all collections and generates vector
embeddings. Re-run after adding new documents or collections.
### 4. Verify
```bash
qmd status # shows index health, collection stats, model info
```
## Search Patterns
### Fast Keyword Search (BM25)
Best for: exact terms, code identifiers, names, known phrases.
No models loaded — near-instant results.
```bash
qmd search "authentication middleware"
qmd search "handleError async"
```
### Semantic Vector Search
Best for: natural language questions, conceptual queries.
Loads embedding model (~3s first query).
```bash
qmd vsearch "how does the rate limiter handle burst traffic"
qmd vsearch "ideas for improving onboarding flow"
```
### Hybrid Search with Reranking (Best Quality)
Best for: important queries where quality matters most.
Uses all 3 models — query expansion, parallel BM25+vector, reranking.
```bash
qmd query "what decisions were made about the database migration"
```
### Structured Multi-Mode Queries
Combine different search types in a single query for precision:
```bash
# BM25 for exact term + vector for concept
qmd query $'lex: rate limiter\nvec: how does throttling work under load'
# With query expansion
qmd query $'expand: database migration plan\nlex: "schema change"'
```
### Query Syntax (lex/BM25 mode)
| Syntax | Effect | Example |
|--------|--------|---------|
| `term` | Prefix match | `perf` matches "performance" |
| `"phrase"` | Exact phrase | `"rate limiter"` |
| `-term` | Exclude term | `performance -sports` |
### HyDE (Hypothetical Document Embeddings)
For complex topics, write what you expect the answer to look like:
```bash
qmd query $'hyde: The migration plan involves three phases. First, we add the new columns without dropping the old ones. Then we backfill data. Finally we cut over and remove legacy columns.'
```
### Scoping to Collections
```bash
qmd search "query" --collection notes
qmd query "query" --collection project-docs
```
### Output Formats
```bash
qmd search "query" --json # JSON output (best for parsing)
qmd search "query" --limit 5 # Limit results
qmd get "#abc123" # Get by document ID
qmd get "path/to/file.md" # Get by file path
qmd get "file.md:50" -l 100 # Get specific line range
qmd multi-get "journals/*.md" --json # Batch retrieve by glob
```
## MCP Integration (Recommended)
qmd exposes an MCP server that provides search tools directly to
Hermes Agent via the native MCP client. This is the preferred
integration — once configured, the agent gets qmd tools automatically
without needing to load this skill.
### Option A: Stdio Mode (Simple)
Add to `~/.hermes/config.yaml`:
```yaml
mcp_servers:
qmd:
command: "qmd"
args: ["mcp"]
timeout: 30
connect_timeout: 45
```
This registers tools: `mcp_qmd_search`, `mcp_qmd_vsearch`,
`mcp_qmd_deep_search`, `mcp_qmd_get`, `mcp_qmd_status`.
**Tradeoff:** Models load on first search call (~19s cold start),
then stay warm for the session. Acceptable for occasional use.
### Option B: HTTP Daemon Mode (Fast, Recommended for Heavy Use)
Start the qmd daemon separately — it keeps models warm in memory:
```bash
# Start daemon (persists across agent restarts)
qmd mcp --http --daemon
# Runs on http://localhost:8181 by default
```
Then configure Hermes Agent to connect via HTTP:
```yaml
mcp_servers:
qmd:
url: "http://localhost:8181/mcp"
timeout: 30
```
**Tradeoff:** Uses ~2GB RAM while running, but every query is fast
(~2-3s). Best for users who search frequently.
### Keeping the Daemon Running
#### macOS (launchd)
```bash
cat > ~/Library/LaunchAgents/com.qmd.daemon.plist << 'EOF'
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN"
"http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
<key>Label</key>
<string>com.qmd.daemon</string>
<key>ProgramArguments</key>
<array>
<string>qmd</string>
<string>mcp</string>
<string>--http</string>
<string>--daemon</string>
</array>
<key>RunAtLoad</key>
<true/>
<key>KeepAlive</key>
<true/>
<key>StandardOutPath</key>
<string>/tmp/qmd-daemon.log</string>
<key>StandardErrorPath</key>
<string>/tmp/qmd-daemon.log</string>
</dict>
</plist>
EOF
launchctl load ~/Library/LaunchAgents/com.qmd.daemon.plist
```
#### Linux (systemd user service)
```bash
mkdir -p ~/.config/systemd/user
cat > ~/.config/systemd/user/qmd-daemon.service << 'EOF'
[Unit]
Description=QMD MCP Daemon
After=network.target
[Service]
ExecStart=qmd mcp --http --daemon
Restart=on-failure
RestartSec=10
Environment=PATH=/usr/local/bin:/usr/bin:/bin
[Install]
WantedBy=default.target
EOF
systemctl --user daemon-reload
systemctl --user enable --now qmd-daemon
systemctl --user status qmd-daemon
```
### MCP Tools Reference
Once connected, these tools are available as `mcp_qmd_*`:
| MCP Tool | Maps To | Description |
|----------|---------|-------------|
| `mcp_qmd_search` | `qmd search` | BM25 keyword search |
| `mcp_qmd_vsearch` | `qmd vsearch` | Semantic vector search |
| `mcp_qmd_deep_search` | `qmd query` | Hybrid search + reranking |
| `mcp_qmd_get` | `qmd get` | Retrieve document by ID or path |
| `mcp_qmd_status` | `qmd status` | Index health and stats |
The MCP tools accept structured JSON queries for multi-mode search:
```json
{
"searches": [
{"type": "lex", "query": "authentication middleware"},
{"type": "vec", "query": "how user login is verified"}
],
"collections": ["project-docs"],
"limit": 10
}
```
## CLI Usage (Without MCP)
When MCP is not configured, use qmd directly via terminal:
```
terminal(command="qmd query 'what was decided about the API redesign' --json", timeout=30)
```
For setup and management tasks, always use terminal:
```
terminal(command="qmd collection add ~/Documents/notes --name notes")
terminal(command="qmd context add qmd://notes 'Personal research notes and ideas'")
terminal(command="qmd embed")
terminal(command="qmd status")
```
## How the Search Pipeline Works
Understanding the internals helps choose the right search mode:
1. **Query Expansion** — A fine-tuned 1.7B model generates 2 alternative
queries. The original gets 2x weight in fusion.
2. **Parallel Retrieval** — BM25 (SQLite FTS5) and vector search run
simultaneously across all query variants.
3. **RRF Fusion** — Reciprocal Rank Fusion (k=60) merges results.
Top-rank bonus: #1 gets +0.05, #2-3 get +0.02.
4. **LLM Reranking** — qwen3-reranker scores top 30 candidates (0.0-1.0).
5. **Position-Aware Blending** — Ranks 1-3: 75% retrieval / 25% reranker.
Ranks 4-10: 60/40. Ranks 11+: 40/60 (trusts reranker more for long tail).
**Smart Chunking:** Documents are split at natural break points (headings,
code blocks, blank lines) targeting ~900 tokens with 15% overlap. Code
blocks are never split mid-block.
## Best Practices
1. **Always add context descriptions** — `qmd context add` dramatically
improves retrieval accuracy. Describe what each collection contains.
2. **Re-embed after adding documents** — `qmd embed` must be re-run when
new files are added to collections.
3. **Use `qmd search` for speed** — when you need fast keyword lookup
(code identifiers, exact names), BM25 is instant and needs no models.
4. **Use `qmd query` for quality** — when the question is conceptual or
the user needs the best possible results, use hybrid search.
5. **Prefer MCP integration** — once configured, the agent gets native
tools without needing to load this skill each time.
6. **Daemon mode for frequent users** — if the user searches their
knowledge base regularly, recommend the HTTP daemon setup.
7. **First query in structured search gets 2x weight** — put the most
important/certain query first when combining lex and vec.
## Troubleshooting
### "Models downloading on first run"
Normal — qmd auto-downloads ~2GB of GGUF models on first use.
This is a one-time operation.
### Cold start latency (~19s)
This happens when models aren't loaded in memory. Solutions:
- Use HTTP daemon mode (`qmd mcp --http --daemon`) to keep warm
- Use `qmd search` (BM25 only) when models aren't needed
- MCP stdio mode loads models on first search, stays warm for session
### macOS: "unable to load extension"
Install Homebrew SQLite: `brew install sqlite`
Then ensure it's on PATH before system SQLite.
### "No collections found"
Run `qmd collection add <path> --name <name>` to add directories,
then `qmd embed` to index them.
### Embedding model override (CJK/multilingual)
Set `QMD_EMBED_MODEL` environment variable for non-English content:
```bash
export QMD_EMBED_MODEL="your-multilingual-model"
```
## Data Storage
- **Index & vectors:** `~/.cache/qmd/index.sqlite`
- **Models:** Auto-downloaded to local cache on first run
- **No cloud dependencies** — everything runs locally
## References
- [GitHub: tobi/qmd](https://github.com/tobi/qmd)
- [QMD Changelog](https://github.com/tobi/qmd/blob/main/CHANGELOG.md)

350
website/docs/user-guide/skills/optional/research/research-scrapling.md Normal file
View file

@ -0,0 +1,350 @@
---
title: "Scrapling"
sidebar_label: "Scrapling"
description: "Web scraping with Scrapling - HTTP fetching, stealth browser automation, Cloudflare bypass, and spider crawling via CLI and Python"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Scrapling
Web scraping with Scrapling - HTTP fetching, stealth browser automation, Cloudflare bypass, and spider crawling via CLI and Python.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/research/scrapling` |
| Path | `optional-skills/research/scrapling` |
| Version | `1.0.0` |
| Author | FEUAZUR |
| License | MIT |
| Tags | `Web Scraping`, `Browser`, `Cloudflare`, `Stealth`, `Crawling`, `Spider` |
| Related skills | [`duckduckgo-search`](/docs/user-guide/skills/optional/research/research-duckduckgo-search), [`domain-intel`](/docs/user-guide/skills/optional/research/research-domain-intel) |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Scrapling
[Scrapling](https://github.com/D4Vinci/Scrapling) is a web scraping framework with anti-bot bypass, stealth browser automation, and a spider framework. It provides three fetching strategies (HTTP, dynamic JS, stealth/Cloudflare) and a full CLI.
**This skill is for educational and research purposes only.** Users must comply with local/international data scraping laws and respect website Terms of Service.
## When to Use
- Scraping static HTML pages (faster than browser tools)
- Scraping JS-rendered pages that need a real browser
- Bypassing Cloudflare Turnstile or bot detection
- Crawling multiple pages with a spider
- When the built-in `web_extract` tool does not return the data you need
## Installation
```bash
pip install "scrapling[all]"
scrapling install
```
Minimal install (HTTP only, no browser):
```bash
pip install scrapling
```
With browser automation only:
```bash
pip install "scrapling[fetchers]"
scrapling install
```
## Quick Reference
| Approach | Class | Use When |
|----------|-------|----------|
| HTTP | `Fetcher` / `FetcherSession` | Static pages, APIs, fast bulk requests |
| Dynamic | `DynamicFetcher` / `DynamicSession` | JS-rendered content, SPAs |
| Stealth | `StealthyFetcher` / `StealthySession` | Cloudflare, anti-bot protected sites |
| Spider | `Spider` | Multi-page crawling with link following |
## CLI Usage
### Extract Static Page
```bash
scrapling extract get 'https://example.com' output.md
```
With CSS selector and browser impersonation:
```bash
scrapling extract get 'https://example.com' output.md \
--css-selector '.content' \
--impersonate 'chrome'
```
### Extract JS-Rendered Page
```bash
scrapling extract fetch 'https://example.com' output.md \
--css-selector '.dynamic-content' \
--disable-resources \
--network-idle
```
### Extract Cloudflare-Protected Page
```bash
scrapling extract stealthy-fetch 'https://protected-site.com' output.html \
--solve-cloudflare \
--block-webrtc \
--hide-canvas
```
### POST Request
```bash
scrapling extract post 'https://example.com/api' output.json \
--json '{"query": "search term"}'
```
### Output Formats
The output format is determined by the file extension:
- `.html` -- raw HTML
- `.md` -- converted to Markdown
- `.txt` -- plain text
- `.json` / `.jsonl` -- JSON
## Python: HTTP Scraping
### Single Request
```python
from scrapling.fetchers import Fetcher
page = Fetcher.get('https://quotes.toscrape.com/')
quotes = page.css('.quote .text::text').getall()
for q in quotes:
print(q)
```
### Session (Persistent Cookies)
```python
from scrapling.fetchers import FetcherSession
with FetcherSession(impersonate='chrome') as session:
page = session.get('https://example.com/', stealthy_headers=True)
links = page.css('a::attr(href)').getall()
for link in links[:5]:
sub = session.get(link)
print(sub.css('h1::text').get())
```
### POST / PUT / DELETE
```python
page = Fetcher.post('https://api.example.com/data', json={"key": "value"})
page = Fetcher.put('https://api.example.com/item/1', data={"name": "updated"})
page = Fetcher.delete('https://api.example.com/item/1')
```
### With Proxy
```python
page = Fetcher.get('https://example.com', proxy='http://user:pass@proxy:8080')
```
## Python: Dynamic Pages (JS-Rendered)
For pages that require JavaScript execution (SPAs, lazy-loaded content):
```python
from scrapling.fetchers import DynamicFetcher
page = DynamicFetcher.fetch('https://example.com', headless=True)
data = page.css('.js-loaded-content::text').getall()
```
### Wait for Specific Element
```python
page = DynamicFetcher.fetch(
'https://example.com',
wait_selector=('.results', 'visible'),
network_idle=True,
)
```
### Disable Resources for Speed
Blocks fonts, images, media, stylesheets (~25% faster):
```python
from scrapling.fetchers import DynamicSession
with DynamicSession(headless=True, disable_resources=True, network_idle=True) as session:
page = session.fetch('https://example.com')
items = page.css('.item::text').getall()
```
### Custom Page Automation
```python
from playwright.sync_api import Page
from scrapling.fetchers import DynamicFetcher
def scroll_and_click(page: Page):
page.mouse.wheel(0, 3000)
page.wait_for_timeout(1000)
page.click('button.load-more')
page.wait_for_selector('.extra-results')
page = DynamicFetcher.fetch('https://example.com', page_action=scroll_and_click)
results = page.css('.extra-results .item::text').getall()
```
## Python: Stealth Mode (Anti-Bot Bypass)
For Cloudflare-protected or heavily fingerprinted sites:
```python
from scrapling.fetchers import StealthyFetcher
page = StealthyFetcher.fetch(
'https://protected-site.com',
headless=True,
solve_cloudflare=True,
block_webrtc=True,
hide_canvas=True,
)
content = page.css('.protected-content::text').getall()
```
### Stealth Session
```python
from scrapling.fetchers import StealthySession
with StealthySession(headless=True, solve_cloudflare=True) as session:
page1 = session.fetch('https://protected-site.com/page1')
page2 = session.fetch('https://protected-site.com/page2')
```
## Element Selection
All fetchers return a `Selector` object with these methods:
### CSS Selectors
```python
page.css('h1::text').get() # First h1 text
page.css('a::attr(href)').getall() # All link hrefs
page.css('.quote .text::text').getall() # Nested selection
```
### XPath
```python
page.xpath('//div[@class="content"]/text()').getall()
page.xpath('//a/@href').getall()
```
### Find Methods
```python
page.find_all('div', class_='quote') # By tag + attribute
page.find_by_text('Read more', tag='a') # By text content
page.find_by_regex(r'\$\d+\.\d{2}') # By regex pattern
```
### Similar Elements
Find elements with similar structure (useful for product listings, etc.):
```python
first_product = page.css('.product')[0]
all_similar = first_product.find_similar()
```
### Navigation
```python
el = page.css('.target')[0]
el.parent # Parent element
el.children # Child elements
el.next_sibling # Next sibling
el.prev_sibling # Previous sibling
```
## Python: Spider Framework
For multi-page crawling with link following:
```python
from scrapling.spiders import Spider, Request, Response
class QuotesSpider(Spider):
name = "quotes"
start_urls = ["https://quotes.toscrape.com/"]
concurrent_requests = 10
download_delay = 1
async def parse(self, response: Response):
for quote in response.css('.quote'):
yield {
"text": quote.css('.text::text').get(),
"author": quote.css('.author::text').get(),
"tags": quote.css('.tag::text').getall(),
}
next_page = response.css('.next a::attr(href)').get()
if next_page:
yield response.follow(next_page)
result = QuotesSpider().start()
print(f"Scraped {len(result.items)} quotes")
result.items.to_json("quotes.json")
```
### Multi-Session Spider
Route requests to different fetcher types:
```python
from scrapling.fetchers import FetcherSession, AsyncStealthySession
class SmartSpider(Spider):
name = "smart"
start_urls = ["https://example.com/"]
def configure_sessions(self, manager):
manager.add("fast", FetcherSession(impersonate="chrome"))
manager.add("stealth", AsyncStealthySession(headless=True), lazy=True)
async def parse(self, response: Response):
for link in response.css('a::attr(href)').getall():
if "protected" in link:
yield Request(link, sid="stealth")
else:
yield Request(link, sid="fast", callback=self.parse)
```
### Pause/Resume Crawling
```python
spider = QuotesSpider(crawldir="./crawl_checkpoint")
spider.start() # Ctrl+C to pause, re-run to resume from checkpoint
```
## Pitfalls
- **Browser install required**: run `scrapling install` after pip install -- without it, `DynamicFetcher` and `StealthyFetcher` will fail
- **Timeouts**: DynamicFetcher/StealthyFetcher timeout is in **milliseconds** (default 30000), Fetcher timeout is in **seconds**
- **Cloudflare bypass**: `solve_cloudflare=True` adds 5-15 seconds to fetch time -- only enable when needed
- **Resource usage**: StealthyFetcher runs a real browser -- limit concurrent usage
- **Legal**: always check robots.txt and website ToS before scraping. This library is for educational and research purposes
- **Python version**: requires Python 3.10+

172
website/docs/user-guide/skills/optional/security/security-1password.md Normal file
View file

@ -0,0 +1,172 @@
---
title: "1Password — Set up and use 1Password CLI (op)"
sidebar_label: "1Password"
description: "Set up and use 1Password CLI (op)"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# 1Password
Set up and use 1Password CLI (op). Use when installing the CLI, enabling desktop app integration, signing in, and reading/injecting secrets for commands.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/security/1password` |
| Path | `optional-skills/security/1password` |
| Version | `1.0.0` |
| Author | arceus77-7, enhanced by Hermes Agent |
| License | MIT |
| Tags | `security`, `secrets`, `1password`, `op`, `cli` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# 1Password CLI
Use this skill when the user wants secrets managed through 1Password instead of plaintext env vars or files.
## Requirements
- 1Password account
- 1Password CLI (`op`) installed
- One of: desktop app integration, service account token (`OP_SERVICE_ACCOUNT_TOKEN`), or Connect server
- `tmux` available for stable authenticated sessions during Hermes terminal calls (desktop app flow only)
## When to Use
- Install or configure 1Password CLI
- Sign in with `op signin`
- Read secret references like `op://Vault/Item/field`
- Inject secrets into config/templates using `op inject`
- Run commands with secret env vars via `op run`
## Authentication Methods
### Service Account (recommended for Hermes)
Set `OP_SERVICE_ACCOUNT_TOKEN` in `~/.hermes/.env` (the skill will prompt for this on first load).
No desktop app needed. Supports `op read`, `op inject`, `op run`.
```bash
export OP_SERVICE_ACCOUNT_TOKEN="your-token-here"
op whoami # verify — should show Type: SERVICE_ACCOUNT
```
### Desktop App Integration (interactive)
1. Enable in 1Password desktop app: Settings → Developer → Integrate with 1Password CLI
2. Ensure app is unlocked
3. Run `op signin` and approve the biometric prompt
### Connect Server (self-hosted)
```bash
export OP_CONNECT_HOST="http://localhost:8080"
export OP_CONNECT_TOKEN="your-connect-token"
```
## Setup
1. Install CLI:
```bash
# macOS
brew install 1password-cli
# Linux (official package/install docs)
# See references/get-started.md for distro-specific links.
# Windows (winget)
winget install AgileBits.1Password.CLI
```
2. Verify:
```bash
op --version
```
3. Choose an auth method above and configure it.
## Hermes Execution Pattern (desktop app flow)
Hermes terminal commands are non-interactive by default and can lose auth context between calls.
For reliable `op` use with desktop app integration, run sign-in and secret operations inside a dedicated tmux session.
Note: This is NOT needed when using `OP_SERVICE_ACCOUNT_TOKEN` — the token persists across terminal calls automatically.
```bash
SOCKET_DIR="${TMPDIR:-/tmp}/hermes-tmux-sockets"
mkdir -p "$SOCKET_DIR"
SOCKET="$SOCKET_DIR/hermes-op.sock"
SESSION="op-auth-$(date +%Y%m%d-%H%M%S)"
tmux -S "$SOCKET" new -d -s "$SESSION" -n shell
# Sign in (approve in desktop app when prompted)
tmux -S "$SOCKET" send-keys -t "$SESSION":0.0 -- "eval \"\$(op signin --account my.1password.com)\"" Enter
# Verify auth
tmux -S "$SOCKET" send-keys -t "$SESSION":0.0 -- "op whoami" Enter
# Example read
tmux -S "$SOCKET" send-keys -t "$SESSION":0.0 -- "op read 'op://Private/Npmjs/one-time password?attribute=otp'" Enter
# Capture output when needed
tmux -S "$SOCKET" capture-pane -p -J -t "$SESSION":0.0 -S -200
# Cleanup
tmux -S "$SOCKET" kill-session -t "$SESSION"
```
## Common Operations
### Read a secret
```bash
op read "op://app-prod/db/password"
```
### Get OTP
```bash
op read "op://app-prod/npm/one-time password?attribute=otp"
```
### Inject into template
```bash
echo "db_password: {{ op://app-prod/db/password }}" | op inject
```
### Run a command with secret env var
```bash
export DB_PASSWORD="op://app-prod/db/password"
op run -- sh -c '[ -n "$DB_PASSWORD" ] && echo "DB_PASSWORD is set" || echo "DB_PASSWORD missing"'
```
## Guardrails
- Never print raw secrets back to user unless they explicitly request the value.
- Prefer `op run` / `op inject` instead of writing secrets into files.
- If command fails with "account is not signed in", run `op signin` again in the same tmux session.
- If desktop app integration is unavailable (headless/CI), use service account token flow.
## CI / Headless note
For non-interactive use, authenticate with `OP_SERVICE_ACCOUNT_TOKEN` and avoid interactive `op signin`.
Service accounts require CLI v2.18.0+.
## References
- `references/get-started.md`
- `references/cli-examples.md`
- https://developer.1password.com/docs/cli/
- https://developer.1password.com/docs/service-accounts/

424
website/docs/user-guide/skills/optional/security/security-oss-forensics.md Normal file
View file

@ -0,0 +1,424 @@
---
title: "Oss Forensics — Supply chain investigation, evidence recovery, and forensic analysis for GitHub repositories"
sidebar_label: "Oss Forensics"
description: "Supply chain investigation, evidence recovery, and forensic analysis for GitHub repositories"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Oss Forensics
Supply chain investigation, evidence recovery, and forensic analysis for GitHub repositories.
Covers deleted commit recovery, force-push detection, IOC extraction, multi-source evidence
collection, hypothesis formation/validation, and structured forensic reporting.
Inspired by RAPTOR's 1800+ line OSS Forensics system.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/security/oss-forensics` |
| Path | `optional-skills/security/oss-forensics` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# OSS Security Forensics Skill
A 7-phase multi-agent investigation framework for researching open-source supply chain attacks.
Adapted from RAPTOR's forensics system. Covers GitHub Archive, Wayback Machine, GitHub API,
local git analysis, IOC extraction, evidence-backed hypothesis formation and validation,
and final forensic report generation.
---
## ⚠️ Anti-Hallucination Guardrails
Read these before every investigation step. Violating them invalidates the report.
1. **Evidence-First Rule**: Every claim in any report, hypothesis, or summary MUST cite at least one evidence ID (`EV-XXXX`). Assertions without citations are forbidden.
2. **STAY IN YOUR LANE**: Each sub-agent (investigator) has a single data source. Do NOT mix sources. The GH Archive investigator does not query the GitHub API, and vice versa. Role boundaries are hard.
3. **Fact vs. Hypothesis Separation**: Mark all unverified inferences with `[HYPOTHESIS]`. Only statements verified against original sources may be stated as facts.
4. **No Evidence Fabrication**: The hypothesis validator MUST mechanically check that every cited evidence ID actually exists in the evidence store before accepting a hypothesis.
5. **Proof-Required Disproval**: A hypothesis cannot be dismissed without a specific, evidence-backed counter-argument. "No evidence found" is not sufficient to disprove—it only makes a hypothesis inconclusive.
6. **SHA/URL Double-Verification**: Any commit SHA, URL, or external identifier cited as evidence must be independently confirmed from at least two sources before being marked as verified.
7. **Suspicious Code Rule**: Never run code found inside the investigated repository locally. Analyze statically only, or use `execute_code` in a sandboxed environment.
8. **Secret Redaction**: Any API keys, tokens, or credentials discovered during investigation must be redacted in the final report. Log them internally only.
---
## Example Scenarios
- **Scenario A: Dependency Confusion**: A malicious package `internal-lib-v2` is uploaded to NPM with a higher version than the internal one. The investigator must track when this package was first seen and if any PushEvents in the target repo updated `package.json` to this version.
- **Scenario B: Maintainer Takeover**: A long-term contributor's account is used to push a backdoored `.github/workflows/build.yml`. The investigator looks for PushEvents from this user after a long period of inactivity or from a new IP/location (if detectable via BigQuery).
- **Scenario C: Force-Push Hide**: A developer accidentally commits a production secret, then force-pushes to "fix" it. The investigator uses `git fsck` and GH Archive to recover the original commit SHA and verify what was leaked.
---
> **Path convention**: Throughout this skill, `SKILL_DIR` refers to the root of this skill's
> installation directory (the folder containing this `SKILL.md`). When the skill is loaded,
> resolve `SKILL_DIR` to the actual path — e.g. `~/.hermes/skills/security/oss-forensics/`
> or the `optional-skills/` equivalent. All script and template references are relative to it.
## Phase 0: Initialization
1. Create investigation working directory:
```bash
mkdir investigation_$(echo "REPO_NAME" | tr '/' '_')
cd investigation_$(echo "REPO_NAME" | tr '/' '_')
```
2. Initialize the evidence store:
```bash
python3 SKILL_DIR/scripts/evidence-store.py --store evidence.json list
```
3. Copy the forensic report template:
```bash
cp SKILL_DIR/templates/forensic-report.md ./investigation-report.md
```
4. Create an `iocs.md` file to track Indicators of Compromise as they are discovered.
5. Record the investigation start time, target repository, and stated investigation goal.
---
## Phase 1: Prompt Parsing and IOC Extraction
**Goal**: Extract all structured investigative targets from the user's request.
**Actions**:
- Parse the user prompt and extract:
- Target repository (`owner/repo`)
- Target actors (GitHub handles, email addresses)
- Time window of interest (commit date ranges, PR timestamps)
- Provided Indicators of Compromise: commit SHAs, file paths, package names, IP addresses, domains, API keys/tokens, malicious URLs
- Any linked vendor security reports or blog posts
**Tools**: Reasoning only, or `execute_code` for regex extraction from large text blocks.
**Output**: Populate `iocs.md` with extracted IOCs. Each IOC must have:
- Type (from: COMMIT_SHA, FILE_PATH, API_KEY, SECRET, IP_ADDRESS, DOMAIN, PACKAGE_NAME, ACTOR_USERNAME, MALICIOUS_URL, OTHER)
- Value
- Source (user-provided, inferred)
**Reference**: See [evidence-types.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/security/oss-forensics/references/evidence-types.md) for IOC taxonomy.
---
## Phase 2: Parallel Evidence Collection
Spawn up to 5 specialist investigator sub-agents using `delegate_task` (batch mode, max 3 concurrent). Each investigator has a **single data source** and must not mix sources.
> **Orchestrator note**: Pass the IOC list from Phase 1 and the investigation time window in the `context` field of each delegated task.
---
### Investigator 1: Local Git Investigator
**ROLE BOUNDARY**: You query the LOCAL GIT REPOSITORY ONLY. Do not call any external APIs.
**Actions**:
```bash
# Clone repository
git clone https://github.com/OWNER/REPO.git target_repo && cd target_repo
# Full commit log with stats
git log --all --full-history --stat --format="%H|%ae|%an|%ai|%s" > ../git_log.txt
# Detect force-push evidence (orphaned/dangling commits)
git fsck --lost-found --unreachable 2>&1 | grep commit > ../dangling_commits.txt
# Check reflog for rewritten history
git reflog --all > ../reflog.txt
# List ALL branches including deleted remote refs
git branch -a -v > ../branches.txt
# Find suspicious large binary additions
git log --all --diff-filter=A --name-only --format="%H %ai" -- "*.so" "*.dll" "*.exe" "*.bin" > ../binary_additions.txt
# Check for GPG signature anomalies
git log --show-signature --format="%H %ai %aN" > ../signature_check.txt 2>&1
```
**Evidence to collect** (add via `python3 SKILL_DIR/scripts/evidence-store.py add`):
- Each dangling commit SHA → type: `git`
- Force-push evidence (reflog showing history rewrite) → type: `git`
- Unsigned commits from verified contributors → type: `git`
- Suspicious binary file additions → type: `git`
**Reference**: See [recovery-techniques.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/security/oss-forensics/references/recovery-techniques.md) for accessing force-pushed commits.
---
### Investigator 2: GitHub API Investigator
**ROLE BOUNDARY**: You query the GITHUB REST API ONLY. Do not run git commands locally.
**Actions**:
```bash
# Commits (paginated)
curl -s "https://api.github.com/repos/OWNER/REPO/commits?per_page=100" > api_commits.json
# Pull Requests including closed/deleted
curl -s "https://api.github.com/repos/OWNER/REPO/pulls?state=all&per_page=100" > api_prs.json
# Issues
curl -s "https://api.github.com/repos/OWNER/REPO/issues?state=all&per_page=100" > api_issues.json
# Contributors and collaborator changes
curl -s "https://api.github.com/repos/OWNER/REPO/contributors" > api_contributors.json
# Repository events (last 300)
curl -s "https://api.github.com/repos/OWNER/REPO/events?per_page=100" > api_events.json
# Check specific suspicious commit SHA details
curl -s "https://api.github.com/repos/OWNER/REPO/git/commits/SHA" > commit_detail.json
# Releases
curl -s "https://api.github.com/repos/OWNER/REPO/releases?per_page=100" > api_releases.json
# Check if a specific commit exists (force-pushed commits may 404 on commits/ but succeed on git/commits/)
curl -s "https://api.github.com/repos/OWNER/REPO/commits/SHA" | jq .sha
```
**Cross-reference targets** (flag discrepancies as evidence):
- PR exists in archive but missing from API → evidence of deletion
- Contributor in archive events but not in contributors list → evidence of permission revocation
- Commit in archive PushEvents but not in API commit list → evidence of force-push/deletion
**Reference**: See [evidence-types.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/security/oss-forensics/references/evidence-types.md) for GH event types.
---
### Investigator 3: Wayback Machine Investigator
**ROLE BOUNDARY**: You query the WAYBACK MACHINE CDX API ONLY. Do not use the GitHub API.
**Goal**: Recover deleted GitHub pages (READMEs, issues, PRs, releases, wiki pages).
**Actions**:
```bash
# Search for archived snapshots of the repo main page
curl -s "https://web.archive.org/cdx/search/cdx?url=github.com/OWNER/REPO&output=json&limit=100&from=YYYYMMDD&to=YYYYMMDD" > wayback_main.json
# Search for a specific deleted issue
curl -s "https://web.archive.org/cdx/search/cdx?url=github.com/OWNER/REPO/issues/NUM&output=json&limit=50" > wayback_issue_NUM.json
# Search for a specific deleted PR
curl -s "https://web.archive.org/cdx/search/cdx?url=github.com/OWNER/REPO/pull/NUM&output=json&limit=50" > wayback_pr_NUM.json
# Fetch the best snapshot of a page
# Use the Wayback Machine URL: https://web.archive.org/web/TIMESTAMP/ORIGINAL_URL
# Example: https://web.archive.org/web/20240101000000*/github.com/OWNER/REPO
# Advanced: Search for deleted releases/tags
curl -s "https://web.archive.org/cdx/search/cdx?url=github.com/OWNER/REPO/releases/tag/*&output=json" > wayback_tags.json
# Advanced: Search for historical wiki changes
curl -s "https://web.archive.org/cdx/search/cdx?url=github.com/OWNER/REPO/wiki/*&output=json" > wayback_wiki.json
```
**Evidence to collect**:
- Archived snapshots of deleted issues/PRs with their content
- Historical README versions showing changes
- Evidence of content present in archive but missing from current GitHub state
**Reference**: See [github-archive-guide.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/security/oss-forensics/references/github-archive-guide.md) for CDX API parameters.
---
### Investigator 4: GH Archive / BigQuery Investigator
**ROLE BOUNDARY**: You query GITHUB ARCHIVE via BIGQUERY ONLY. This is a tamper-proof record of all public GitHub events.
> **Prerequisites**: Requires Google Cloud credentials with BigQuery access (`gcloud auth application-default login`). If unavailable, skip this investigator and note it in the report.
**Cost Optimization Rules** (MANDATORY):
1. ALWAYS run a `--dry_run` before every query to estimate cost.
2. Use `_TABLE_SUFFIX` to filter by date range and minimize scanned data.
3. Only SELECT the columns you need.
4. Add a LIMIT unless aggregating.
```bash
# Template: safe BigQuery query for PushEvents to OWNER/REPO
bq query --use_legacy_sql=false --dry_run "
SELECT created_at, actor.login, payload.commits, payload.before, payload.head,
payload.size, payload.distinct_size
FROM \`githubarchive.month.*\`
WHERE _TABLE_SUFFIX BETWEEN 'YYYYMM' AND 'YYYYMM'
AND type = 'PushEvent'
AND repo.name = 'OWNER/REPO'
LIMIT 1000
"
# If cost is acceptable, re-run without --dry_run
# Detect force-pushes: zero-distinct_size PushEvents mean commits were force-erased
# payload.distinct_size = 0 AND payload.size > 0 → force push indicator
# Check for deleted branch events
bq query --use_legacy_sql=false "
SELECT created_at, actor.login, payload.ref, payload.ref_type
FROM \`githubarchive.month.*\`
WHERE _TABLE_SUFFIX BETWEEN 'YYYYMM' AND 'YYYYMM'
AND type = 'DeleteEvent'
AND repo.name = 'OWNER/REPO'
LIMIT 200
"
```
**Evidence to collect**:
- Force-push events (payload.size > 0, payload.distinct_size = 0)
- DeleteEvents for branches/tags
- WorkflowRunEvents for suspicious CI/CD automation
- PushEvents that precede a "gap" in the git log (evidence of rewrite)
**Reference**: See [github-archive-guide.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/security/oss-forensics/references/github-archive-guide.md) for all 12 event types and query patterns.
---
### Investigator 5: IOC Enrichment Investigator
**ROLE BOUNDARY**: You enrich EXISTING IOCs from Phase 1 using passive public sources ONLY. Do not execute any code from the target repository.
**Actions**:
- For each commit SHA: attempt recovery via direct GitHub URL (`github.com/OWNER/REPO/commit/SHA.patch`)
- For each domain/IP: check passive DNS, WHOIS records (via `web_extract` on public WHOIS services)
- For each package name: check npm/PyPI for matching malicious package reports
- For each actor username: check GitHub profile, contribution history, account age
- Recover force-pushed commits using 3 methods (see [recovery-techniques.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/security/oss-forensics/references/recovery-techniques.md))
---
## Phase 3: Evidence Consolidation
After all investigators complete:
1. Run `python3 SKILL_DIR/scripts/evidence-store.py --store evidence.json list` to see all collected evidence.
2. For each piece of evidence, verify the `content_sha256` hash matches the original source.
3. Group evidence by:
- **Timeline**: Sort all timestamped evidence chronologically
- **Actor**: Group by GitHub handle or email
- **IOC**: Link evidence to the IOC it relates to
4. Identify **discrepancies**: items present in one source but absent in another (key deletion indicators).
5. Flag evidence as `[VERIFIED]` (confirmed from 2+ independent sources) or `[UNVERIFIED]` (single source only).
---
## Phase 4: Hypothesis Formation
A hypothesis must:
- State a specific claim (e.g., "Actor X force-pushed to BRANCH on DATE to erase commit SHA")
- Cite at least 2 evidence IDs that support it (`EV-XXXX`, `EV-YYYY`)
- Identify what evidence would disprove it
- Be labeled `[HYPOTHESIS]` until validated
**Common hypothesis templates** (see [investigation-templates.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/security/oss-forensics/references/investigation-templates.md)):
- Maintainer Compromise: legitimate account used post-takeover to inject malicious code
- Dependency Confusion: package name squatting to intercept installs
- CI/CD Injection: malicious workflow changes to run code during builds
- Typosquatting: near-identical package name targeting misspellers
- Credential Leak: token/key accidentally committed then force-pushed to erase
For each hypothesis, spawn a `delegate_task` sub-agent to attempt to find disconfirming evidence before confirming.
---
## Phase 5: Hypothesis Validation
The validator sub-agent MUST mechanically check:
1. For each hypothesis, extract all cited evidence IDs.
2. Verify each ID exists in `evidence.json` (hard failure if any ID is missing → hypothesis rejected as potentially fabricated).
3. Verify each `[VERIFIED]` piece of evidence was confirmed from 2+ sources.
4. Check logical consistency: does the timeline depicted by the evidence support the hypothesis?
5. Check for alternative explanations: could the same evidence pattern arise from a benign cause?
**Output**:
- `VALIDATED`: All evidence cited, verified, logically consistent, no plausible alternative explanation.
- `INCONCLUSIVE`: Evidence supports hypothesis but alternative explanations exist or evidence is insufficient.
- `REJECTED`: Missing evidence IDs, unverified evidence cited as fact, logical inconsistency detected.
Rejected hypotheses feed back into Phase 4 for refinement (max 3 iterations).
---
## Phase 6: Final Report Generation
Populate `investigation-report.md` using the template in [forensic-report.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/security/oss-forensics/templates/forensic-report.md).
**Mandatory sections**:
- Executive Summary: one-paragraph verdict (Compromised / Clean / Inconclusive) with confidence level
- Timeline: chronological reconstruction of all significant events with evidence citations
- Validated Hypotheses: each with status and supporting evidence IDs
- Evidence Registry: table of all `EV-XXXX` entries with source, type, and verification status
- IOC List: all extracted and enriched Indicators of Compromise
- Chain of Custody: how evidence was collected, from what sources, at what timestamps
- Recommendations: immediate mitigations if compromise detected; monitoring recommendations
**Report rules**:
- Every factual claim must have at least one `[EV-XXXX]` citation
- Executive Summary must state confidence level (High / Medium / Low)
- All secrets/credentials must be redacted to `[REDACTED]`
---
## Phase 7: Completion
1. Run final evidence count: `python3 SKILL_DIR/scripts/evidence-store.py --store evidence.json list`
2. Archive the full investigation directory.
3. If compromise is confirmed:
- List immediate mitigations (rotate credentials, pin dependency hashes, notify affected users)
- Identify affected versions/packages
- Note disclosure obligations (if a public package: coordinate with the package registry)
4. Present the final `investigation-report.md` to the user.
---
## Ethical Use Guidelines
This skill is designed for **defensive security investigation** — protecting open-source software from supply chain attacks. It must not be used for:
- **Harassment or stalking** of contributors or maintainers
- **Doxing** — correlating GitHub activity to real identities for malicious purposes
- **Competitive intelligence** — investigating proprietary or internal repositories without authorization
- **False accusations** — publishing investigation results without validated evidence (see anti-hallucination guardrails)
Investigations should be conducted with the principle of **minimal intrusion**: collect only the evidence necessary to validate or refute the hypothesis. When publishing results, follow responsible disclosure practices and coordinate with affected maintainers before public disclosure.
If the investigation reveals a genuine compromise, follow the coordinated vulnerability disclosure process:
1. Notify the repository maintainers privately first
2. Allow reasonable time for remediation (typically 90 days)
3. Coordinate with package registries (npm, PyPI, etc.) if published packages are affected
4. File a CVE if appropriate
---
## API Rate Limiting
GitHub REST API enforces rate limits that will interrupt large investigations if not managed.
**Authenticated requests**: 5,000/hour (requires `GITHUB_TOKEN` env var or `gh` CLI auth)
**Unauthenticated requests**: 60/hour (unusable for investigations)
**Best practices**:
- Always authenticate: `export GITHUB_TOKEN=ghp_...` or use `gh` CLI (auto-authenticates)
- Use conditional requests (`If-None-Match` / `If-Modified-Since` headers) to avoid consuming quota on unchanged data
- For paginated endpoints, fetch all pages in sequence — don't parallelize against the same endpoint
- Check `X-RateLimit-Remaining` header; if below 100, pause for `X-RateLimit-Reset` timestamp
- BigQuery has its own quotas (10 TiB/day free tier) — always dry-run first
- Wayback Machine CDX API: no formal rate limit, but be courteous (1-2 req/sec max)
If rate-limited mid-investigation, record the partial results in the evidence store and note the limitation in the report.
---
## Reference Materials
- [github-archive-guide.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/security/oss-forensics/references/github-archive-guide.md) — BigQuery queries, CDX API, 12 event types
- [evidence-types.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/security/oss-forensics/references/evidence-types.md) — IOC taxonomy, evidence source types, observation types
- [recovery-techniques.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/security/oss-forensics/references/recovery-techniques.md) — Recovering deleted commits, PRs, issues
- [investigation-templates.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/security/oss-forensics/references/investigation-templates.md) — Pre-built hypothesis templates per attack type
- [evidence-store.py](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/security/oss-forensics/scripts/evidence-store.py) — CLI tool for managing the evidence JSON store
- [forensic-report.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/security/oss-forensics/templates/forensic-report.md) — Structured report template

207
website/docs/user-guide/skills/optional/security/security-sherlock.md Normal file
View file

@ -0,0 +1,207 @@
---
title: "Sherlock — OSINT username search across 400+ social networks"
sidebar_label: "Sherlock"
description: "OSINT username search across 400+ social networks"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Sherlock
OSINT username search across 400+ social networks. Hunt down social media accounts by username.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/security/sherlock` |
| Path | `optional-skills/security/sherlock` |
| Version | `1.0.0` |
| Author | unmodeled-tyler |
| License | MIT |
| Tags | `osint`, `security`, `username`, `social-media`, `reconnaissance` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Sherlock OSINT Username Search
Hunt down social media accounts by username across 400+ social networks using the [Sherlock Project](https://github.com/sherlock-project/sherlock).
## When to Use
- User asks to find accounts associated with a username
- User wants to check username availability across platforms
- User is conducting OSINT or reconnaissance research
- User asks "where is this username registered?" or similar
## Requirements
- Sherlock CLI installed: `pipx install sherlock-project` or `pip install sherlock-project`
- Alternatively: Docker available (`docker run -it --rm sherlock/sherlock`)
- Network access to query social platforms
## Procedure
### 1. Check if Sherlock is Installed
**Before doing anything else**, verify sherlock is available:
```bash
sherlock --version
```
If the command fails:
- Offer to install: `pipx install sherlock-project` (recommended) or `pip install sherlock-project`
- **Do NOT** try multiple installation methods — pick one and proceed
- If installation fails, inform the user and stop
### 2. Extract Username
**Extract the username directly from the user's message if clearly stated.**
Examples where you should **NOT** use clarify:
- "Find accounts for nasa" → username is `nasa`
- "Search for johndoe123" → username is `johndoe123`
- "Check if alice exists on social media" → username is `alice`
- "Look up user bob on social networks" → username is `bob`
**Only use clarify if:**
- Multiple potential usernames mentioned ("search for alice or bob")
- Ambiguous phrasing ("search for my username" without specifying)
- No username mentioned at all ("do an OSINT search")
When extracting, take the **exact** username as stated — preserve case, numbers, underscores, etc.
### 3. Build Command
**Default command** (use this unless user specifically requests otherwise):
```bash
sherlock --print-found --no-color "<username>" --timeout 90
```
**Optional flags** (only add if user explicitly requests):
- `--nsfw` — Include NSFW sites (only if user asks)
- `--tor` — Route through Tor (only if user asks for anonymity)
**Do NOT ask about options via clarify** — just run the default search. Users can request specific options if needed.
### 4. Execute Search
Run via the `terminal` tool. The command typically takes 30-120 seconds depending on network conditions and site count.
**Example terminal call:**
```json
{
"command": "sherlock --print-found --no-color \"target_username\"",
"timeout": 180
}
```
### 5. Parse and Present Results
Sherlock outputs found accounts in a simple format. Parse the output and present:
1. **Summary line:** "Found X accounts for username 'Y'"
2. **Categorized links:** Group by platform type if helpful (social, professional, forums, etc.)
3. **Output file location:** Sherlock saves results to `<username>.txt` by default
**Example output parsing:**
```
[+] Instagram: https://instagram.com/username
[+] Twitter: https://twitter.com/username
[+] GitHub: https://github.com/username
```
Present findings as clickable links when possible.
## Pitfalls
### No Results Found
If Sherlock finds no accounts, this is often correct — the username may not be registered on checked platforms. Suggest:
- Checking spelling/variation
- Trying similar usernames with `?` wildcard: `sherlock "user?name"`
- The user may have privacy settings or deleted accounts
### Timeout Issues
Some sites are slow or block automated requests. Use `--timeout 120` to increase wait time, or `--site` to limit scope.
### Tor Configuration
`--tor` requires Tor daemon running. If user wants anonymity but Tor isn't available, suggest:
- Installing Tor service
- Using `--proxy` with an alternative proxy
### False Positives
Some sites always return "found" due to their response structure. Cross-reference unexpected results with manual checks.
### Rate Limiting
Aggressive searches may trigger rate limits. For bulk username searches, add delays between calls or use `--local` with cached data.
## Installation
### pipx (recommended)
```bash
pipx install sherlock-project
```
### pip
```bash
pip install sherlock-project
```
### Docker
```bash
docker pull sherlock/sherlock
docker run -it --rm sherlock/sherlock <username>
```
### Linux packages
Available on Debian 13+, Ubuntu 22.10+, Homebrew, Kali, BlackArch.
## Ethical Use
This tool is for legitimate OSINT and research purposes only. Remind users:
- Only search usernames they own or have permission to investigate
- Respect platform terms of service
- Do not use for harassment, stalking, or illegal activities
- Consider privacy implications before sharing results
## Verification
After running sherlock, verify:
1. Output lists found sites with URLs
2. `<username>.txt` file created (default output) if using file output
3. If `--print-found` used, output should only contain `[+]` lines for matches
## Example Interaction
**User:** "Can you check if the username 'johndoe123' exists on social media?"
**Agent procedure:**
1. Check `sherlock --version` (verify installed)
2. Username provided — proceed directly
3. Run: `sherlock --print-found --no-color "johndoe123" --timeout 90`
4. Parse output and present links
**Response format:**
> Found 12 accounts for username 'johndoe123':
>
> • https://twitter.com/johndoe123
> • https://github.com/johndoe123
> • https://instagram.com/johndoe123
> • [... additional links]
>
> Results saved to: johndoe123.txt
---
**User:** "Search for username 'alice' including NSFW sites"
**Agent procedure:**
1. Check sherlock installed
2. Username + NSFW flag both provided
3. Run: `sherlock --print-found --no-color --nsfw "alice" --timeout 90`
4. Present results

206
website/docs/user-guide/skills/optional/web-development/web-development-page-agent.md Normal file
View file

@ -0,0 +1,206 @@
---
title: "Page Agent"
sidebar_label: "Page Agent"
description: "Embed alibaba/page-agent into your own web application — a pure-JavaScript in-page GUI agent that ships as a single <script> tag or npm package and lets end-..."
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Page Agent
Embed alibaba/page-agent into your own web application — a pure-JavaScript in-page GUI agent that ships as a single &lt;script> tag or npm package and lets end-users of your site drive the UI with natural language ("click login, fill username as John"). No Python, no headless browser, no extension required. Use this skill when the user is a web developer who wants to add an AI copilot to their SaaS / admin panel / B2B tool, make a legacy web app accessible via natural language, or evaluate page-agent against a local (Ollama) or cloud (Qwen / OpenAI / OpenRouter) LLM. NOT for server-side browser automation — point those users to Hermes' built-in browser tool instead.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/web-development/page-agent` |
| Path | `optional-skills/web-development/page-agent` |
| Version | `1.0.0` |
| Author | Hermes Agent |
| License | MIT |
| Tags | `web`, `javascript`, `agent`, `browser`, `gui`, `alibaba`, `embed`, `copilot`, `saas` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# page-agent
alibaba/page-agent (https://github.com/alibaba/page-agent, 17k+ stars, MIT) is an in-page GUI agent written in TypeScript. It lives inside a webpage, reads the DOM as text (no screenshots, no multi-modal LLM), and executes natural-language instructions like "click the login button, then fill username as John" against the current page. Pure client-side — the host site just includes a script and passes an OpenAI-compatible LLM endpoint.
## When to use this skill
Load this skill when a user wants to:
- **Ship an AI copilot inside their own web app** (SaaS, admin panel, B2B tool, ERP, CRM) — "users on my dashboard should be able to type 'create invoice for Acme Corp and email it' instead of clicking through five screens"
- **Modernize a legacy web app** without rewriting the frontend — page-agent drops on top of existing DOM
- **Add accessibility via natural language** — voice / screen-reader users drive the UI by describing what they want
- **Demo or evaluate page-agent** against a local (Ollama) or hosted (Qwen, OpenAI, OpenRouter) LLM
- **Build interactive training / product demos** — let an AI walk a user through "how to submit an expense report" live in the real UI
## When NOT to use this skill
- User wants **Hermes itself to drive a browser** → use Hermes' built-in browser tool (Browserbase / Camofox). page-agent is the *opposite* direction.
- User wants **cross-tab automation without embedding** → use Playwright, browser-use, or the page-agent Chrome extension
- User needs **visual grounding / screenshots** → page-agent is text-DOM only; use a multimodal browser agent instead
## Prerequisites
- Node 22.13+ or 24+, npm 10+ (docs claim 11+ but 10.9 works fine)
- An OpenAI-compatible LLM endpoint: Qwen (DashScope), OpenAI, Ollama, OpenRouter, or anything speaking `/v1/chat/completions`
- Browser with devtools (for debugging)
## Path 1 — 30-second demo via CDN (no install)
Fastest way to see it work. Uses alibaba's free testing LLM proxy — **for evaluation only**, subject to their terms.
Add to any HTML page (or paste into the devtools console as a bookmarklet):
```html
<script src="https://cdn.jsdelivr.net/npm/page-agent@1.8.0/dist/iife/page-agent.demo.js" crossorigin="true"></script>
```
A panel appears. Type an instruction. Done.
Bookmarklet form (drop into bookmarks bar, click on any page):
```javascript
javascript:(function(){var s=document.createElement('script');s.src='https://cdn.jsdelivr.net/npm/page-agent@1.8.0/dist/iife/page-agent.demo.js';document.head.appendChild(s);})();
```
## Path 2 — npm install into your own web app (production use)
Inside an existing web project (React / Vue / Svelte / plain):
```bash
npm install page-agent
```
Wire it up with your own LLM endpoint — **never ship the demo CDN to real users**:
```javascript
import { PageAgent } from 'page-agent'
const agent = new PageAgent({
model: 'qwen3.5-plus',
baseURL: 'https://dashscope.aliyuncs.com/compatible-mode/v1',
apiKey: process.env.LLM_API_KEY, // never hardcode
language: 'en-US',
})
// Show the panel for end users:
agent.panel.show()
// Or drive it programmatically:
await agent.execute('Click submit button, then fill username as John')
```
Provider examples (any OpenAI-compatible endpoint works):
| Provider | `baseURL` | `model` |
|----------|-----------|---------|
| Qwen / DashScope | `https://dashscope.aliyuncs.com/compatible-mode/v1` | `qwen3.5-plus` |
| OpenAI | `https://api.openai.com/v1` | `gpt-4o-mini` |
| Ollama (local) | `http://localhost:11434/v1` | `qwen3:14b` |
| OpenRouter | `https://openrouter.ai/api/v1` | `anthropic/claude-sonnet-4.6` |
**Key config fields** (passed to `new PageAgent({...})`):
- `model`, `baseURL`, `apiKey` — LLM connection
- `language` — UI language (`en-US`, `zh-CN`, etc.)
- Allowlist and data-masking hooks exist for locking down what the agent can touch — see https://alibaba.github.io/page-agent/ for the full option list
**Security.** Don't put your `apiKey` in client-side code for a real deployment — proxy LLM calls through your backend and point `baseURL` at your proxy. The demo CDN exists because alibaba runs that proxy for evaluation.
## Path 3 — clone the source repo (contributing, or hacking on it)
Use this when the user wants to modify page-agent itself, test it against arbitrary sites via a local IIFE bundle, or develop the browser extension.
```bash
git clone https://github.com/alibaba/page-agent.git
cd page-agent
npm ci # exact lockfile install (or `npm i` to allow updates)
```
Create `.env` in the repo root with an LLM endpoint. Example:
```
LLM_MODEL_NAME=gpt-4o-mini
LLM_API_KEY=sk-...
LLM_BASE_URL=https://api.openai.com/v1
```
Ollama flavor:
```
LLM_BASE_URL=http://localhost:11434/v1
LLM_API_KEY=NA
LLM_MODEL_NAME=qwen3:14b
```
Common commands:
```bash
npm start # docs/website dev server
npm run build # build every package
npm run dev:demo # serve IIFE bundle at http://localhost:5174/page-agent.demo.js
npm run dev:ext # develop the browser extension (WXT + React)
npm run build:ext # build the extension
```
**Test on any website** using the local IIFE bundle. Add this bookmarklet:
```javascript
javascript:(function(){var s=document.createElement('script');s.src=`http://localhost:5174/page-agent.demo.js?t=${Math.random()}`;s.onload=()=>console.log('PageAgent ready!');document.head.appendChild(s);})();
```
Then: `npm run dev:demo`, click the bookmarklet on any page, and the local build injects. Auto-rebuilds on save.
**Warning:** your `.env` `LLM_API_KEY` is inlined into the IIFE bundle during dev builds. Don't share the bundle. Don't commit it. Don't paste the URL into Slack. (Verified: grepping the public dev bundle returns the literal values from `.env`.)
## Repo layout (Path 3)
Monorepo with npm workspaces. Key packages:
| Package | Path | Purpose |
|---------|------|---------|
| `page-agent` | `packages/page-agent/` | Main entry with UI panel |
| `@page-agent/core` | `packages/core/` | Core agent logic, no UI |
| `@page-agent/mcp` | `packages/mcp/` | MCP server (beta) |
| — | `packages/llms/` | LLM client |
| — | `packages/page-controller/` | DOM ops + visual feedback |
| — | `packages/ui/` | Panel + i18n |
| — | `packages/extension/` | Chrome/Firefox extension |
| — | `packages/website/` | Docs + landing site |
## Verifying it works
After Path 1 or Path 2:
1. Open the page in a browser with devtools open
2. You should see a floating panel. If not, check the console for errors (most common: CORS on the LLM endpoint, wrong `baseURL`, or a bad API key)
3. Type a simple instruction matching something visible on the page ("click the Login link")
4. Watch the Network tab — you should see a request to your `baseURL`
After Path 3:
1. `npm run dev:demo` prints `Accepting connections at http://localhost:5174`
2. `curl -I http://localhost:5174/page-agent.demo.js` returns `HTTP/1.1 200 OK` with `Content-Type: application/javascript`
3. Click the bookmarklet on any site; panel appears
## Pitfalls
- **Demo CDN in production** — don't. It's rate-limited, uses alibaba's free proxy, and their terms forbid production use.
- **API key exposure** — any key passed to `new PageAgent({apiKey: ...})` ships in your JS bundle. Always proxy through your own backend for real deployments.
- **Non-OpenAI-compatible endpoints** fail silently or with cryptic errors. If your provider needs native Anthropic/Gemini formatting, use an OpenAI-compatibility proxy (LiteLLM, OpenRouter) in front.
- **CSP blocks** — sites with strict Content-Security-Policy may refuse to load the CDN script or disallow inline eval. In that case, self-host from your origin.
- **Restart dev server** after editing `.env` in Path 3 — Vite only reads env at startup.
- **Node version** — the repo declares `^22.13.0 || >=24`. Node 20 will fail `npm ci` with engine errors.
- **npm 10 vs 11** — docs say npm 11+; npm 10.9 actually works fine.
## Reference
- Repo: https://github.com/alibaba/page-agent
- Docs: https://alibaba.github.io/page-agent/
- License: MIT (built on browser-use's DOM processing internals, Copyright 2024 Gregor Zunic)