Tools And Tool Calling

Integrator defines a tool function with no type hints and a one-line docstring, then wonders why the model passes wrong-typed arguments.

Decorate with @tool and provide type hints + a clear docstring; LangChain derives the tool name, description, and args schema from them. The description is what the model uses to route, so it must describe purpose and arguments precisely.

The model returns two tool_calls in one AIMessage; the integrator returns a single merged ToolMessage without matching tool_call_id.

For each AIMessage.tool_calls entry, execute the tool and return one ToolMessage whose tool_call_id matches the call's id. The next model turn requires every tool_call to have a matching ToolMessage; mismatched or missing ids break the conversation contract.

Two tools, get_weather (current) and get_forecast (future), have near-identical descriptions, so the model routes 'will it rain tomorrow?' to the wrong tool.

Write distinct, scope-bounding descriptions: get_weather 'current conditions only; for future dates use get_forecast'. The model routes from descriptions, so disambiguation must live there, not in a system-prompt tie-breaker that biases against one tool.