Developers / Docs / JSON output

JSON output

Use response_format to ask for a JSON answer, optionally shaped by a JSON Schema. ARK instructs the model to answer in JSON and strips any Markdown fence around it.

API reference: Chat completions

Any JSON object

response = client.chat.completions.create(
    model="arklabsai/Qwen3.6-35B-A3B",
    messages=[
        {"role": "system", "content": "Extract invoice data."},
        {"role": "user", "content": "Invoice FV/2026/10/118 from Odra Logistics, 12 400 PLN net, due 30 October 2026."},
    ],
    response_format={"type": "json_object"},
)
data = json.loads(response.choices[0].message.content)

JSON that follows a schema

Pass a schema with json_schema. name is required; description helps the model.

Python
response = client.chat.completions.create(
    model="arklabsai/Qwen3.5-122B-A10B",
    messages=[{"role": "user", "content": "Invoice FV/2026/10/118 from Odra Logistics, 12 400 PLN net, due 30 October 2026."}],
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "invoice",
            "schema": {
                "type": "object",
                "properties": {
                    "number": {"type": "string"},
                    "vendor": {"type": "string"},
                    "net_amount": {"type": "number"},
                    "currency": {"type": "string"},
                    "due_date": {"type": "string", "format": "date"},
                },
                "required": ["number", "vendor", "net_amount", "currency", "due_date"],
            },
        },
    },
)
Validate what you receive

ARK guides the model towards JSON but does not constrain decoding, and strict: true is not enforced. Parse the answer, validate it against your schema, and retry on the rare failure. Keep max_tokens high enough that the object is not cut off (finish_reason: "length").

With the Anthropic API

On the Messages API, use output_config.format: {"type": "json_schema", "schema": {…}}. It behaves the same way.