Structured outputs & JSON mode

Force a model to return valid JSON. MeshTok supports two modes from the OpenAI spec: json_object (loose JSON) and json_schema (strict schema enforcement on supporting models). All 86 JSON-capable models support at least json_object.

Mode 1: json_object (works everywhere)

Sets response_format to JSON. You must include the word "JSON" in the prompt or the request will be rejected.

from openai import OpenAI
client = OpenAI(base_url="https://meshtok.com/v1", api_key="sk-...")

resp = client.chat.completions.create(
    model="deepseek/deepseek-v4-flash",
    messages=[{"role": "user", "content": "List 3 fruits as JSON: {name, color}."}],
    response_format={"type": "json_object"},
)
import json
data = json.loads(resp.choices[0].message.content)

Mode 2: json_schema (strict)

Constrains the model to a specific JSON schema — every field, every type, every enum enforced. Supported by GPT-5.x, Claude 5, GLM 5.2, Gemini 2.5, and Qwen 3.

resp = client.chat.completions.create(
    model="bigmodel/glm-5.2",
    messages=[{"role": "user", "content": "Extract the company from: Apple announced..."}],
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "company",
            "strict": True,
            "schema": {
                "type": "object",
                "properties": {
                    "name": {"type": "string"},
                    "ticker": {"type": "string"},
                    "founded": {"type": "integer"}
                },
                "required": ["name", "ticker", "founded"],
                "additionalProperties": False,
            }
        }
    },
)

Tips for reliable JSON

• Use json_schema over json_object whenever the model supports it — it eliminates malformed responses.
• Lower temperature (0.0–0.3) for extraction tasks; the schema does the rest.
• For streaming, parse each delta and reconstruct — schema-constrained streams emit field-by-field.
• If a model rejects strict: true, fall back to json_object plus Pydantic validation.