Function / Tool Calling#

Tool calling is the bridge between LLMs and external systems (APIs, databases, the OS).

Schema Design#

LLMs need well-defined JSON schemas to understand what tools are available.

tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Get current weather in a location",
        "parameters": {
            "type": "object",
            "properties": {"location": {"type": "string"}}
        }
    }
}]

Parallel Tool Calling#

Modern LLMs can invoke multiple tools simultaneously. If asked “What is the weather in Tokyo and Paris?”, the model can return two tool calls in a single response, dramatically reducing latency.

Best Practices#

• Descriptions matter: The LLM relies heavily on parameter descriptions.
• Graceful failure: If a tool fails, return the error to the LLM so it can correct its parameters.