Issue with Structured Output in OpenAI API

API
1

Hi everyone,

I wanted to share an issue we encountered while trying to implement Structured Outputs in our application. My goal was to enforce a specific JSON schema using the new structured output features described in the docs.

According to the documentation, Structured Outputs should be enabled by supplying a text property with a format object. Our payload looked like this:

text: {
  format: {
    type: "json_schema",
    name: "tool_schema",
    schema: {
      type: "object",
      properties: {
        title: {
          type: "string",
          description: "Main title of the tool."
        },
        coverSlide: {
          type: "object",
          properties: {
            title: {
              type: "string",
              description: "Cover slide title."
            },
            subtitle: {
              type: "string",
              description: "Cover slide subtitle."
            }
          },
          required: ["title", "subtitle"],
          additionalProperties: false
        },
        // ... rest of schema ...
      },
      required: [
        "title",
        "coverSlide",
        "problemSlide",
        "solutionSlide",
        "researchSlide",
        "commonProblemsSlide",
        "stepsSlide",
        "ctaSlide"
      ],
      additionalProperties: false
    }
  }
}

I used this with the chat/completions endpoint and tried it on models such as "o1-2024-12-17", “gpt-4o-2024-08-06”`, “gpt-4o”, and “o1” expecting the model to output JSON strictly adhering to my schema.

{
  "error": {
    "message": "Unrecognized request argument supplied: text",
    "type": "invalid_request_error",
    "param": null,
    "code": null
  }
}

Despite the docs showing an example that includes a text property with json_schema formatting, the API didn’t recognize the text parameter. We double-checked our payload against the documentation, and it was correct according to the examples provided.

After exploring the docs and community feedback, I removed the text property and used the response_format parameter with { type: "json_object" } as a fallback. While this approach doesn’t enforce the schema adherence as strictly as the json_schema format was intended to, it allowed me to receive structured JSON output.

Has anyone else encountered this issue with structured outputs?

2

When you are reading documentation on the documentation page, at the upper-right, there is a selection for either Chat Completions or for the newer Responses endpoint.

In the API documentation, the left bar of endpoints will require discovery of each endpoint in the list.

text → format is the Responses endpoints version of Chat Completions’ response_format parameter.

While each would have the same “schema” ultimately placed, the structure of the metadata object container for signifying the type, name and whether it is strict is also different between endpoints.