Structured outputs
Ask a model for JSON, either in the prompt or with a JSON schema.
You can ask a model to return JSON in two ways: describe the shape you want in
the prompt, or send a JSON schema in response_format.
Both response_format: {"type": "json_schema", ...} and
response_format: {"type": "json_object"} are supported.
Prompt-based JSON
Describe the schema in the system message, then parse the reply. Nothing forces
the model to follow the prompt, so be ready for a json.JSONDecodeError:
import json, os
from openai import OpenAI
client = OpenAI(
base_url="https://blaze.e2enetworks.net/v1",
api_key=os.environ["BLAZE_API_KEY"],
)
response = client.chat.completions.create(
model="qwen/qwen3.8-27b",
messages=[
{
"role": "system",
"content": (
"Reply with JSON only, matching "
'{"title": string, "tags": string[]}. No prose.'
),
},
{"role": "user", "content": "A short article about vector databases."},
],
)
data = json.loads(response.choices[0].message.content)Schema-enforced output
Send the schema in response_format, as with OpenAI:
{
"model": "google/gemma-4-31b-it",
"messages": [
{ "role": "user", "content": "Give a title for an article about vector databases." }
],
"response_format": {
"type": "json_schema",
"json_schema": {
"name": "article",
"strict": true,
"schema": {
"type": "object",
"properties": { "title": { "type": "string" } },
"required": ["title"]
}
}
}
}Two things to watch:
- Give every array a
maxItems. Without a limit, the model can repeat items until it reachesmax_tokens. - Leave room in
max_tokensfor reasoning, because reasoning counts towards it.