9. Function/tool calling fundamentals

Tool calling turns an LLM into a controller. You describe tools as JSON schemas; the model returns a structured call; your code executes it and feeds the result back as a tool message.

ts
const tools = [{
  type: "function",
  function: {
    name: "get_gas_price",
    description: "Current gas price in gwei for an EVM chain",
    parameters: { type: "object", properties: { chainId: { type: "integer" } }, required: ["chainId"], additionalProperties: false },
  },
}];

let messages = [{ role: "user", content: "Is Base cheap right now?" }];
const first = await client.chat.completions.create({ model, messages, tools });
for (const call of first.choices[0].message.tool_calls ?? []) {
  const args = JSON.parse(call.function.arguments);
  const result = await getGasPrice(args.chainId);        // you run it, with your own authz
  messages.push(first.choices[0].message, { role: "tool", tool_call_id: call.id, content: JSON.stringify(result) });
}
const final = await client.chat.completions.create({ model, messages, tools });

Design rules: few tools with sharp descriptions beat many vague ones; validate arguments server-side; make tools idempotent where possible; return compact results (the model pays for every token).

Gotcha: never let the model's arguments decide authorisation. Check permissions in your handler against the real session, not against anything the model said.

Read in another language

How to study this topic

Scoring

Check your understanding

Badges