Tool use

A language model cannot do arithmetic reliably, look up the current weather or read your database. But it can ask to be allowed to: you describe tools (a name, a description, the parameters as a JSON schema), the model answers with a structured call {"name": "weather", "arguments": {"city": "Umeå"}}, your code runs the tool and sends the result back, and the model formulates the answer.

The model proposes; you execute. That means you:

• validate the arguments against the schema (the model can invent fields),
• limit the tool (read-only, a whitelist, a timeout, a budget),
• return errors as text the model can react to, rather than crashing,
• log every call.

A tool that can delete files or send money should require human confirmation.

import json
from pydantic import BaseModel, ValidationError

class WeatherArgs(BaseModel):
    city: str

TOOLS = {
  "weather": {"schema": WeatherArgs, "fn": lambda a: {"temp_c": get_temp(a.city)},
              "description": "The current temperature in a Swedish city"},
}

def tool_spec():
    return [{"type": "function", "function": {"name": n, "description": v["description"],
             "parameters": v["schema"].model_json_schema()}} for n, v in TOOLS.items()]

def run_call(call):
    v = TOOLS.get(call["name"])
    if not v:
        return {"error": f"unknown tool {call['name']}"}
    try:
        args = v["schema"].model_validate(json.loads(call["arguments"]))
        return v["fn"](args)
    except (ValidationError, json.JSONDecodeError) as e:
        return {"error": f"invalid arguments: {e}"}       # the model gets a chance to correct itself
    except Exception as e:
        return {"error": f"the tool failed: {type(e).__name__}"}

# the loop: reply = llm(messages, tools=tool_spec()); if reply.tool_calls: run → append(role="tool") → llm again