Structured output based on JSON Schema

[sebastian-velten]

Hello everyone,

I'm currently migrating from AutoGen and I'm wondering how I can get structured output from my agent based on a given JSON Schema, like I did in the past:

agent = AssistantAgent(
    name=self.config.name,
    system_message=load_system_message(self.config),
    model_client=OpenAIChatCompletionClient(
        model=self.config.alias.model,
        api_key=os.environ["PROVIDER_API_KEY"],
        response_format=response_format or {
            "type": "json_schema",
            "json_schema": {
                "name": self.config.output_schema["title"],
                "strict": True,
                "schema": self.get_output_schema(self.config.output_schema)
            }
        },
        **(await self.getClientProperties())
    ),
    reflect_on_tool_use=False,
    workbench=self.workbenches
)

The JSON Schema is known only at runtime, so it is not possible to model a pydantic class, representing the schema in advance.

[moonbox3]

Hi @sebastian-velten thanks for your question. I just tested this, and notice we need to make a small change to support passing in a JSON schema for response format. That was completed in #2362. You can find a sample on how to do this here.

[sebastian-velten]

Awesome! Thanks a lot for providing a solution :)

I noticed, that the sample code fails when switching from the OpenAIChatClient to the OpenAIResponsesClient. I guess, this approach should work with the responses client, too. Or am I wrong?

[moonbox3]

Hi @sebastian-velten - I agree, I would have thought the responses client would also work; however, upon inspection, I had to adjust some things so that the response_format is properly handled. This will be handled in #2440.

[sebastian-velten]

Great, thanks again