Structured outputs & tools

Tool calling and JSON output, validated by Keln.

Tool calling

Standard OpenAI shape, tools, tool_choice, parallel calls where the model supports them:

{
  "model": "moonshotai/kimi-k2.7-code",
  "messages": [{"role": "user", "content": "What's the weather in San Francisco?"}],
  "tools": [{
    "type": "function",
    "function": {
      "name": "get_weather",
      "parameters": {
        "type": "object",
        "properties": {"city": {"type": "string"}},
        "required": ["city"]
      }
    }
  }]
}

JSON output

Both modes:

• response_format: {"type": "json_object"}, valid JSON, shape up to the model.
• response_format: {"type": "json_schema", "json_schema": {…}}, conform to your schema.

json_object guarantees syntactic JSON and, per the OpenAI contract, expects the word "JSON" to appear somewhere in your prompt. json_schema enforces your shape without that prompt help; prefer it when the shape matters.

Validation

response_format output is checked for parseable JSON (and your schema, where given); tool-call arguments are checked for valid JSON. Non-streaming responses are checked before delivery. Streamed responses are validated as they flow; a route that emits broken structure is penalized and de-routed, so a malformed stream can reach you at most transiently. When the check must complete before delivery, use non-streaming.

Unsupported combinations

When no capacity serving the model supports the requested mode in the transport you asked for, the request returns a 400: response_format_not_supported or tools_not_supported (errors).