Can I Specify a Return Value Schema in OpenAI's Function-Calling Feature?

Below is the current situation regarding defining a function’s return schema:

• The "returns" Field Is Not Officially Supported:
  Although you can include a "returns" field in your function schema, the official documentation and API do not currently use or enforce this field to dictate the output structure of a function call. The API expects you to specify input parameters (via the "parameters" field), not an output or return schema.

• No Method to Directly Specify Return Schema:
  At this time, there isn’t an officially supported method to instruct GPT to format its future output to match a given return schema defined in the function’s schema. The API’s function calling feature interprets the provided "parameters" to understand what input the function expects, but it does not treat any accompanying "returns" field as a directive for how responses should be formatted.

• Alternative Approaches:
  If you need a specified schema for the output of a function call, you may consider:

  • Using the API’s structured outputs capability via the response_format parameter. By setting it to a JSON schema, you can guide the model’s response to follow a defined structure. However, this is separate from the function calling feature and its schema.

  • Handling output validation on your end once you receive the function’s response. You can parse and verify the structure against the desired schema within your application logic.

In summary, while you can include a "returns" field in your function schema, it won’t change the behavior of the model’s response. The GPT function calling mechanism does not currently support making the return schema an active part of how responses are formatted or validated.