Why does the Assistant API always invoke custom functions with 'json_schema' response_format?

Below is an explanation with supporting code formatting:

---

When you configure your assistant with a response format of type json_schema, it tells the API that every response must adhere to your provided JSON schema. This “structured outputs” setting means the model always returns output via your custom function, even for prompts that might not strictly need function invocation (like “How are you?”). This behavior is by design.

If you want the assistant to decide whether to use functions based on the prompt’s context (i.e., only invoke functions when required), you should use the text response format. With text, the model evaluates the prompt and may call functions only when necessary.

Unfortunately, there isn’t a supported configuration for the json_schema mode that allows conditional function usage. You’ll need to decide between:

• Using json_schema: Ensures all output follows the given JSON schema via your defined custom functions.
• Using text: Allows the model to evaluate the context of the prompt and decide if functions are needed.

Your current configuration:

JAVASCRIPTCopied
const assistant = await openai.beta.assistants.create({
  name,
  instructions,
  model: 'gpt-4o-2024-08-06',
  tools: [
    {
      type: 'function',
      function: MY_CUSTOM_FUNCTION,
    },
  ],
  response_format: {
    type: 'json_schema',
    json_schema: {
      name: 'collab_response',
      strict: true,
      schema: collabAISchema,
    },
  },
});

This setup enforces structured outputs, so regardless of the prompt, the assistant will always format its answer using your custom function and JSON schema. If your goal is to have the function be optional, then switching to or incorporating logic with the text response format is advisable.

---

In summary, if you require conditional usage of functions based on the prompt's content, you’d need to use a response format that’s not locked into structured outputs (i.e., text).