To ensure an LLM can invoke a function tool that returns a JSON string and then consume the response, follow this structured approach: • Tool Implementation: Your function tool should return a Python dictionary or list, which will be automatically converted to a JSON string for the LLM to consume. • Response Handling: The system converts the tool's return value to a string. For complex types (like dict or list), it uses

json.dumps()

to serialize the response. If the response is a simple type (str, int, float, bool), it is converted to a string directly. • *Example Tool:*`@llm.ai_callable() async def get_weather(self, location: str): # ... fetch weather data ... return {"weather": "sunny", "temperature_f": 70}` • LLM Consumption: The LLM receives the JSON string and can use it in its reply to the user. Common Errors and Solutions: • JSON Parsing Errors: Ensure the tool returns serializable data. If the tool returns a non-serializable object, use

json.dumps()

or convert it to a string. • Tool Execution Errors: Handle exceptions within the tool and return error messages as strings if needed. • Response Format Issues: Always return a dictionary, list, or string. Avoid returning custom objects unless they are serializable. By following these practices, the LLM will be able to invoke your function tool, receive the JSON response, and use it in user interactions without errors. Sources: Tool definition and use | LiveKit Docs | Function calling with Voice Agents | LiveKit Docs | mcp_client/util.py | common/chat.cpp