When AI output feeds into code (saving to a database, filling a form, calling another API), you need valid JSON every time. Here's how to get there.
1. Ask clearly, with an example
Extract the order details. Reply with JSON only, no other text, in this shape:
{"customer": string, "items": [{"name": string, "qty": number}], "total": number}Showing the exact shape works better than describing it.
2. Use JSON mode where supported
Many models accept response_format:
r = client.chat.completions.create(
model="gpt-5.6-luna",
response_format={"type": "json_object"},
messages=[{"role": "system", "content": "Reply with JSON only."}, {"role": "user", "content": text}],
)Support varies by model and provider. If a model ignores or rejects response_format, rely on the prompt plus validation.
3. Use tool calling as a schema
A trick that works across most models: define a tool whose parameters are the schema you want, and ask the model to call it. The arguments come back as JSON matching your schema.
4. Always validate
import json
from pydantic import BaseModel, ValidationError
class Order(BaseModel):
customer: str
items: list[dict]
total: float
def parse(text: str) -> Order | None:
text = text.strip().removeprefix("```json").removesuffix("```").strip()
try:
return Order.model_validate(json.loads(text))
except (json.JSONDecodeError, ValidationError):
return None5. Retry once with the error
If parsing fails, send the model its own output and the validation error, and ask it to fix the JSON. One retry fixes almost every case.
6. Keep temperature low
For extraction, use temperature: 0 or close to it. Creativity is the enemy of consistent structure.
Claude, GPT, Gemini, DeepSeek, image and video models through one OpenAI- and Anthropic-compatible endpoint. Prepaid, pay only for what you use.