chat.mdx

RAG Debug Mode

Enable debug mode to see detailed retrieval information:

# Python SDK with debug — ChatAgent takes a single ChatAgentConfig
from gaia_agent_chat.agent import ChatAgent, ChatAgentConfig

config = ChatAgentConfig(
    rag_documents=['document.pdf'],
    debug=True,
    silent_mode=False,
)
agent = ChatAgent(config)

result = agent.process_query("What is the vision statement?")
print(result)

Chunking Strategies

```python
config = ChatAgentConfig(
    rag_documents=['document.pdf'],
    chunk_size=500,
    chunk_overlap=50,
)
agent = ChatAgent(config)
```

```python
config = ChatAgentConfig(
    rag_documents=['document.pdf'],
    use_llm_chunking=True,
    chunk_size=500,
)
agent = ChatAgent(config)
```

Troubleshooting

```bash
# Install talk dependencies
uv pip install -e ".[talk]"

# If the error persists, install pip in your environment
python -m ensurepip --upgrade
```

# Better quality
chat.enable_rag(documents=["doc.pdf"], chunk_size=600, max_chunks=5, chunk_overlap=100)

# Memory efficient
chat.enable_rag(documents=["doc.pdf"], chunk_size=400, max_chunks=2)
```

Dynamic Tool Loading

The doc profile can load tools semantically per turn instead of showing the LLM every registered tool on every turn. A small always-on CORE set is combined with tools whose descriptions best match the conversation, which shrinks the first-turn prompt and speeds up the first reply.

It activates only on the doc profile (the registered doc agent, the SDK with ChatAgentConfig(prompt_profile="doc"), or gaia eval agent --agent-type doc). Turn it on with the config field, an environment variable, or the Agent UI Settings → Dynamic Tools (Beta) toggle. The env var wins over both — when it is set, the UI toggle reflects the effective value and disables itself — which is handy for the eval harness:

from gaia_agent_chat.agent import ChatAgent, ChatAgentConfig

agent = ChatAgent(ChatAgentConfig(prompt_profile="doc", dynamic_tools=True))

# Env override (applies wherever a doc-profile ChatAgent runs)
GAIA_DYNAMIC_TOOLS=1 gaia eval agent --category tool_selection --agent-type doc

# Optional tuning: match threshold (cosine, inclusive) and loaded-set cap
GAIA_DYNAMIC_TOOLS_TAU=0.20 GAIA_DYNAMIC_TOOLS_MAX=14 GAIA_DYNAMIC_TOOLS=1 ...

It needs memory enabled (it reuses the memory embedder). If memory is off, the toggle is off, or the embedder is unreachable, the agent automatically falls back to showing all tools — so a turn never loses access to a tool it needs. See the Dynamic Tool Loader plan for the full design.

Next Steps