Structured outputs

response_format constrains what the model writes: valid JSON, or JSON matching a schema you provide. Both shapes follow the OpenAI convention and are forwarded to the serving endpoint.

JSON mode#

{ "response_format": { "type": "json_object" } }

The model must emit a single valid JSON object. As with OpenAI, tell the model what JSON you want in the prompt — JSON mode guarantees well-formedness, not shape.

JSON Schema#

json_schema pins the shape itself:

{
  "model": "deepseek/deepseek-v4-flash-0731-abliterated",
  "messages": [
    { "role": "user", "content": "Extract the order from: two espressos and a croissant, table 5." }
  ],
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "order",
      "strict": true,
      "schema": {
        "type": "object",
        "properties": {
          "table": { "type": "integer" },
          "items": {
            "type": "array",
            "items": {
              "type": "object",
              "properties": {
                "product": { "type": "string" },
                "quantity": { "type": "integer" }
              },
              "required": ["product", "quantity"],
              "additionalProperties": false
            }
          }
        },
        "required": ["table", "items"],
        "additionalProperties": false
      }
    }
  }
}

The assistant message's content is then a JSON string matching the schema:

{
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "{\"table\":5,\"items\":[{\"product\":\"espresso\",\"quantity\":2},{\"product\":\"croissant\",\"quantity\":1}]}"
      },
      "finish_reason": "stop"
    }
  ]
}

Routing#

Support varies by endpoint. An endpoint that honours response_format declares it in its supported_parameters; some additionally declare structured_outputs to signal schema enforcement.

A request setting structured_outputs: true is only ever routed to an endpoint that declares it, with no preference needed — an endpoint that ignores a schema hands you prose where you needed JSON, which is the wrong kind of answer rather than a worse one. response_format alone is also requires an endpoint declaration. The name means different things on different surfaces (JSON mode on chat, the transcript encoding on a transcription). When the shape matters, request structured_outputs so the endpoint's schema compliance check applies:

{
  "response_format": { "type": "json_schema", "json_schema": { "name": "order", "schema": {} } },
  "structured_outputs": true
}

structured_outputs: true in the request body is a capability marker: it does not change what the model generates, it decides where the request may go. Even with strict schemas, parse and validate the output before acting on it — the schema constrains generation, it does not replace validation at your boundary.

The declaration is checked rather than taken. Every endpoint declaring structured_outputs is sent a small schema every six hours, and the check passes only if the reply parses against it — an engine that accepts response_format and answers in prose fails, because that is the failure the declaration exists to rule out. An endpoint whose latest check failed has the promise withheld and stops receiving structured-output traffic; the next passing check, at most half an hour later, restores it. Nothing is required of you: your request routes to an endpoint whose promise holds. A declaration nothing has checked routes normally, since an unverified promise is not a broken one.