Enable JSON mode

Configure JSON mode in Serverless Inference to get structured JSON output from model responses for easier parsing.

Enabling JSON mode instructs the model to return the response in a valid JSON format. However, the response’s schema may not be consistent or adhere to a particular structure. For consistent structured JSON responses, we recommend using structured output when possible.

To enable JSON mode, specify it as the “response_format” in the request:

Python

import json
import openai

client = openai.OpenAI(
    base_url='https://api.inference.wandb.ai/v1',
    api_key="<your-api-key>",  # Create an API key at https://wandb.ai/settings
)

response = client.chat.completions.create(
    model="openai/gpt-oss-20b",
    messages=[
        {"role": "system", "content": "You are a helpful assistant that outputs JSON."},
        {"role": "user", "content": "Give me a list of three fruits with their colors."},
    ],
    response_format={"type": "json_object"}  # This enables JSON mode
)

content = response.choices[0].message.content
parsed = json.loads(content)
print(parsed)

Bash

curl https://api.inference.wandb.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <your-api-key>" \
  -d '{
    "model": "openai/gpt-oss-20b",
    "messages": [
        {"role": "system", "content": "You are a helpful assistant that outputs JSON."},
        {"role": "user", "content": "Give me a list of three fruits with their colors."},
    ],
    "response_format": {"type": "json_object"}
  }'