Function Calling

What it is#

You declare the available tools up front:

{
  "name": "get_weather",
  "description": "Look up current weather for a given city",
  "parameters": {
    "type": "object",
    "properties": {
      "city": { "type": "string", "description": "City name in Latin or local script" }
    },
    "required": ["city"]
  }
}

When the model decides to call one, instead of speaking it emits:

{
  "tool_calls": [
    { "name": "get_weather", "arguments": { "city": "Beijing" } }
  ]
}

Your program sees that JSON, actually calls get_weather("Beijing"), sends "26°C, cloudy" back to the model, and the model phrases the answer for the user.

Analogy#

Key concepts#

How it works#

The model makes decisions, the runtime executes — this is the foundation of agents.

Practical notes#

• Describe the tool's purpose. "Get weather" is much weaker than "Look up today's weather by city to answer 'should I go out?' style questions" — model tool-selection accuracy hinges almost entirely on this.
• Lock down the parameter schema. Use enum for choices, required for mandatory fields, pattern for formats — the model almost never fills nonsense.
• Return structured errors. Don't throw — return { "error": "city not found" } so the model can adjust args and retry.
• Run in parallel. Modern SDKs support N concurrent tool calls per turn — multiple times faster than serial.
• Don't expose too many tools. Above ~20 tools the model's pick rate falls off a cliff. Classify intent first, then route to a sub-toolkit.

Easy confusions#

Further reading#

• MCP — the "USB-C" of Function Calling
• Skills — bundle tools into "capabilities"
• Code Interpreter — the special "tool" of "run arbitrary code"
• ReAct — the loop that repeatedly invokes tools