Tool calling

Tool calling lets a model request a function that your application executes. KosCompute never executes the function for you.

1. Declare the tool¶

{
  "model": "qwen/qwen3.8-27b",
  "messages": [{"role": "user", "content": "What is the weather in Brno?"}],
  "tools": [{
    "type": "function",
    "function": {
      "name": "get_weather",
      "description": "Return current weather for a city",
      "parameters": {
        "type": "object",
        "properties": {"city": {"type": "string"}},
        "required": ["city"],
        "additionalProperties": false
      }
    }
  }],
  "tool_choice": "auto",
  "max_completion_tokens": 400
}

2. Execute the call¶

Read choices[0].message.tool_calls. Parse each function's arguments JSON, validate it again in your application, apply authorization, and execute the local function. Never treat model-generated arguments as trusted input.

3. Return the result¶

Append the assistant message containing tool_calls, then append a tool message with the matching ID:

{"role":"tool","tool_call_id":"call_abc123","content":"{\"temperature_c\":21}"}

Send the expanded conversation in a second completion request.

Tool choice¶

auto lets the model decide, none prevents a call, required requests a call, and a named function selects one tool. Backend behavior for required can be model-specific. In particular, some routes can guarantee required choice only with one named function; the Responses adapter rejects combinations it cannot guarantee instead of pretending they are equivalent.

Responses uses function_call output items and function_call_output input items joined by call_id. See the Responses guide for that shape.