mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-04-26 01:01:40 +00:00
Update environment configuration and enhance tool definitions
- Modified `.env.example` to set default terminal environment to 'local' and updated Docker, Singularity, and Modal image references to use 'python:3.11-slim'. - Updated `package.json` to include Node.js engine requirements and modified post-install script for better user guidance. - Enhanced `pyproject.toml` to reflect new dependencies and optional dependencies for modal and development environments. - Improved `README.md` with additional setup instructions for Singularity and Node.js dependencies, along with clearer toolset documentation. - Refactored `model_tools.py` to include new tool definitions and ensure consistency across toolsets. - Updated architecture documentation to clarify tool structure and registration processes.
This commit is contained in:
teknium
parent
f8846f85a1
commit
7ea17bb957
8 changed files with 535 additions and 257 deletions
|
|
@ -1,114 +1,121 @@
|
|||
# Message Graph
|
||||
# Message Format & Trajectories
|
||||
|
||||
```mermaid
|
||||
graph TD
|
||||
%% Message nodes
|
||||
SystemMsg["📋 System Message<br/>Role: System<br/>Content: Messages are nodes in a graph"]
|
||||
UserMsg["👤 User Message<br/>Role: User<br/>Content: But messages aren't the only thing in the graph"]
|
||||
subgraph PrevMessages["Previous Messages"]
|
||||
PrevSystemMsg["📋 System Message<br/>Role: System<br/>Content: Edits are kept in the graph as context"]
|
||||
PrevUserMsg["👤 User Message<br/>Role: User<br/>Content: So we can ensure they're immutable while keeping them editable"]
|
||||
end
|
||||
|
||||
%% Chat Response as a subgraph
|
||||
subgraph ChatResponseBox["💬 Chat Response"]
|
||||
ChatMetadata["📊 Metadata<br/>Temp: 1.0<br/>..."]
|
||||
ChatResponseText["📝 Response<br/>Hello, Here's a subagent call: <tool>subagent</tool>"]
|
||||
ChatContent["Content: Hello, Here's a subagent call..."]
|
||||
end
|
||||
|
||||
%% Tool Response as a subgraph
|
||||
subgraph ToolResponseBox["🔧 Tool Response"]
|
||||
subgraph ToolMetadata["📊 Tool Metadata"]
|
||||
ToolMetadataLength["Length: 3"]
|
||||
subgraph ToolChat["💭 Subagent Chat"]
|
||||
SubagentSystem["📋 System<br/>Content: Subagent call received"]
|
||||
SubagentUser["👤 User<br/>Content: Process this request"]
|
||||
SubagentAssistant["🤖 Assistant<br/>Content: Processing..."]
|
||||
SubagentSystem --> SubagentUser
|
||||
SubagentUser --> SubagentAssistant
|
||||
end
|
||||
end
|
||||
ToolContent["Content: Subagent call output"]
|
||||
end
|
||||
|
||||
%% Graph flow connections
|
||||
SystemMsg --> UserMsg
|
||||
PrevSystemMsg --> PrevUserMsg
|
||||
PrevMessages -.-> UserMsg
|
||||
UserMsg --> ChatResponseBox
|
||||
ChatResponseBox --> ToolResponseBox
|
||||
|
||||
class SystemMsg,UserMsg messageNode
|
||||
class ChatResponseBox responseNode
|
||||
class ToolResponseBox responseNode
|
||||
class ChatMetadata,ChatResponseText,ChatContent,ToolMetadata,ToolChat,ToolContent,ToolMetadataLength metadataNode
|
||||
```
|
||||
Hermes Agent uses two message formats: the **API format** for LLM calls and the **trajectory format** for training data export.
|
||||
|
||||
Messages should be a graph (DAG, specifically) of immutable elements.
|
||||
## API Message Format
|
||||
|
||||
## Why immutable elements?
|
||||
We want to train on policy
|
||||
- This means the context cannot change after we call a response.
|
||||
|
||||
## Why a graph?
|
||||
Nodes and connections are a natural way to represent the flow of information in an agent conversation.
|
||||
|
||||
## Will this be annoying to deal with?
|
||||
|
||||
It shouldn't be! While there will be internal stuff that may look ???, for the interface, it should be as simple as your
|
||||
normal context window edits, so `message_history[2]['content'] = my_edit`, but internally we'll deal with the recordkeeping
|
||||
and how this ends up parsing into on policy training data, if requested.
|
||||
|
||||
## Edges
|
||||
|
||||
Edges are the connections between nodes, and there are two types we are concerned with:
|
||||
- **Sequential edges**: These represent the flow of conversation, connecting messages in the order they were sent. For example, a user message followed by an assistant response.
|
||||
- **Parallel edges**: These represent versioning, e.g. edit history, context squishing, etc.
|
||||
We, however, are only concerned about parallel edges when we break the prefix, and ignore any other parallel edges.
|
||||
|
||||
## So what does this look like in practice?
|
||||
Standard OpenAI chat format used during execution:
|
||||
|
||||
```python
|
||||
import copy
|
||||
messages = [
|
||||
# System prompt
|
||||
{"role": "system", "content": "You are a helpful assistant with tools..."},
|
||||
|
||||
# User query
|
||||
{"role": "user", "content": "Search for Python tutorials"},
|
||||
|
||||
# Assistant with tool call
|
||||
{
|
||||
"role": "assistant",
|
||||
"content": None,
|
||||
"tool_calls": [{
|
||||
"id": "call_abc123",
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "web_search",
|
||||
"arguments": "{\"query\": \"Python tutorials\"}"
|
||||
}
|
||||
}]
|
||||
},
|
||||
|
||||
# Tool result
|
||||
{
|
||||
"role": "tool",
|
||||
"tool_call_id": "call_abc123",
|
||||
"content": "{\"results\": [...]}"
|
||||
},
|
||||
|
||||
# Final response
|
||||
{"role": "assistant", "content": "Here's what I found..."}
|
||||
]
|
||||
```
|
||||
|
||||
## Trajectory Format (ShareGPT)
|
||||
|
||||
class MessageGraph:
|
||||
def __init__(self):
|
||||
self.messages = []
|
||||
self.prev_graph = None
|
||||
Exported for training in ShareGPT format:
|
||||
|
||||
def append(self, message):
|
||||
self.messages.append(message)
|
||||
```json
|
||||
{
|
||||
"conversations": [
|
||||
{"from": "system", "value": "You are a helpful assistant..."},
|
||||
{"from": "human", "value": "Search for Python tutorials"},
|
||||
{"from": "gpt", "value": "<tool_call>\n{\"name\": \"web_search\", \"arguments\": {\"query\": \"Python tutorials\"}}\n</tool_call>"},
|
||||
{"from": "tool", "value": "<tool_response>\n{\"results\": [...]}\n</tool_response>"},
|
||||
{"from": "gpt", "value": "Here's what I found..."}
|
||||
],
|
||||
"tools": "[{\"type\": \"function\", \"function\": {...}}]",
|
||||
"source": "hermes-agent"
|
||||
}
|
||||
```
|
||||
|
||||
def __getitem__(self, index):
|
||||
return self.messages[index]
|
||||
## Reasoning Content
|
||||
|
||||
def __setitem__(self, key, value):
|
||||
# check if an assistant message is after this indx
|
||||
needs_new_graph = False
|
||||
first_idx = -1
|
||||
for i in range(key, len(self.messages)):
|
||||
if (i == key) and (value['role'] == 'assistant') and (value['content'] == self.messages[i]['content']):
|
||||
# no op
|
||||
return
|
||||
needs_new_graph = needs_new_graph or (self.messages[i]['role'] == 'assistant')
|
||||
if needs_new_graph and first_idx == -1:
|
||||
first_idx = i
|
||||
if needs_new_graph:
|
||||
self.prev_graph = copy.deepcopy(self)
|
||||
self.messages[key] = value
|
||||
For models that output reasoning/chain-of-thought:
|
||||
|
||||
def __len__(self):
|
||||
return len(self.messages)
|
||||
**During execution** (API format):
|
||||
```python
|
||||
# Stored internally but not sent back to model in content
|
||||
assistant_msg = {
|
||||
"role": "assistant",
|
||||
"content": "Here's what I found...",
|
||||
"reasoning": "Let me think about this step by step..." # Internal only
|
||||
}
|
||||
```
|
||||
|
||||
def __eq__(self, other):
|
||||
return "\n\n".join(f"{msg['role']}: {msg['content']}" for msg in self) == "\n\n".join(
|
||||
f"{msg['role']}: {msg['content']}" for msg in other)
|
||||
**In trajectory export** (reasoning wrapped in tags):
|
||||
```json
|
||||
{
|
||||
"from": "gpt",
|
||||
"value": "<think>\nLet me think about this step by step...\n</think>\nHere's what I found..."
|
||||
}
|
||||
```
|
||||
|
||||
## Conversion Flow
|
||||
|
||||
# in use
|
||||
messages = MessageGraph()
|
||||
messages.append({'role': 'system', 'content': 'Hello, I am a system message'})
|
||||
messages[0] = {'role': 'user', 'content': 'Hello, I am a user message'}
|
||||
```
|
||||
```
|
||||
API Response → Internal Storage → Trajectory Export
|
||||
↓ ↓ ↓
|
||||
tool_calls reasoning field <tool_call> tags
|
||||
reasoning_content <think> tags
|
||||
```
|
||||
|
||||
The conversion happens in `_convert_to_trajectory_format()` in `run_agent.py`.
|
||||
|
||||
## Ephemeral System Prompts
|
||||
|
||||
Batch processing supports ephemeral system prompts that guide behavior during execution but are NOT saved to trajectories:
|
||||
|
||||
```python
|
||||
# During execution: full system prompt + ephemeral guidance
|
||||
messages = [
|
||||
{"role": "system", "content": SYSTEM_PROMPT + "\n\n" + ephemeral_prompt},
|
||||
...
|
||||
]
|
||||
|
||||
# In saved trajectory: only the base system prompt
|
||||
trajectory = {
|
||||
"conversations": [
|
||||
{"from": "system", "value": SYSTEM_PROMPT}, # No ephemeral
|
||||
...
|
||||
]
|
||||
}
|
||||
```
|
||||
|
||||
## Trajectory Compression
|
||||
|
||||
Long trajectories can be compressed for training using `trajectory_compressor.py`:
|
||||
|
||||
- Protects first/last N turns
|
||||
- Summarizes middle turns with LLM
|
||||
- Targets specific token budget
|
||||
- See `configs/trajectory_compression.yaml` for settings
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue