How can I enforce JSON-only responses using the OpenAI Assistant API?

Below is some guidance on forcing a JSON response from your assistant using the assistant API:

Use Explicit JSON Instructions

Unlike the completions API’s function_call capability, the assistant API relies on how you instruct the model. To get valid JSON every time, you must explicitly tell the model (via a system or user message) that its response must be valid JSON. For example, you can add a system message such as:

JAVASCRIPTCopied
{
  "role": "system",
  "content": "Provide all output in strict JSON format with no additional commentary. Your response should be a valid JSON object."
}

Enable JSON Mode

When creating your assistant, use the JSON mode configuration. This is done by setting the required field (e.g., { "type": "json_object" } in your assistant configuration) so that the generated message is expected to be valid JSON. Note that you still must instruct the model to output JSON via its messages.

Example Assistant Creation

Below is an example of how you might configure your assistant using Node.js (based on the documentation):

JAVASCRIPTCopied
await openai.beta.assistants.create({
  name: "JSON Assistant",
  instructions: "You are a bot that responds with valid JSON only. Provide no extra commentary. The response should always be a JSON object.",
  tools: [{ "type": "code_interpreter" }],
  model: "gpt-4o",
  response_format: "auto",
  // This is where you specify JSON mode
  json_mode: { "type": "json_object" }
});

Handling Tool Outputs

When you submit tool outputs, you are already sending JSON strings (as shown by your code). As long as your assistant has been instructed to produce JSON, the assistant’s responses should be a valid JSON string. If you later parse the output in your code, you can safely convert the string back into an object:

JAVASCRIPTCopied
if (messageValue && messageValue.text && messageValue.text.value) {
  // Log the JSON response
  console.log(messageValue.text.value);
  // Emit the parsed JSON object
  socket.emit('message', {
    role: 'ai',
    content: JSON.parse(messageValue.text.value)
  });
}

Summary

• Explicit Instructions: Always tell the model via a system/user message to output JSON.
• JSON Mode: Set your assistant’s configuration (json_mode) to enforce JSON responses.
• Tool Outputs: Continue to handle and emit responses as JSON as you are doing.

Using these steps, you can achieve behavior similar to forcing a specific function call output but tailored to ensure your assistant always returns JSON.