Enable JSON mode

Configure JSON mode in W&B 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"}
  }'