Pydantic Support & Separate Package for Responses API Models

[bbqiu]

Confirm this is a feature request for the Python library and not the underlying OpenAI API.

• This is a feature request for the Python library

Describe the feature or improvement you're requesting

Hey team,

I’m a contributor to MLflow. We’re working to update MLflow’s standard agent authoring interface to be an extension on top of + compatible with the Responses API. However, we don’t want authoring an agent with MLflow to require using the OpenAI SDK, so we need to make a copy of all input/output classes (and maintain this copy) in order to enforce input/output schemas. There also seems to be quite a few input/output fields that look like they’re subject to continuous change (literals with preview models / tool names)

Would these two feature requests be possible?

1. Pydantic classes for all input/output objects.
   a. We could replace the dict compatibility for inputs via a wrapper that will cast into Pydantic models. Might help save on maintenance cost of two copies of a lot of classes ex. ResponseOutputMessageParam and ResponseOutputMessage classes, one TypedDict and the other BaseModel.
2. Publish a separate lightweight python package specifically for the Responses API Pydantic classes, so other OSS packages can easily build on top of them

Additional context

No response

[RobertCraigie]

Thanks for the request but I don't think we'll want to publish a separate package just for the types, that would be too much maintenance burden. If you really do not want to add this library as a dependency then I'd recommend coming up with a build script to copy the types and make them work for your use case.

Pydantic classes for all input/output objects.

out of curiosity, would using TypeAdapters be acceptable in your case? we've been trying to avoid duplicating all the params types to reduce naming confusion