Missing default value to logprobs in openai.types.chat.chat_completion.Choice

[davorrunje]

Confirm this is an issue with the Python library and not an underlying OpenAI API

• This is an issue with the Python library

Describe the bug

#980 added token logprobs to chat completions of type Optional[ChoiceLogprobs] in openai.types.chat.chat_completion.Choice and openai.types.chat.chat_completion_chunk.Choice. In the latter, the default value is set to None, while in the former it is not set. This causes backward compatibility problems with code written for versions prior to 1.5.0.

To Reproduce

Execution of the following code fails:

from openai.types.chat.chat_completion import ChatCompletionMessage, Choice

msg = ChatCompletionMessage(role="assistant", content="")

Choice(
    index=0,
    finish_reason="stop",
    message=msg,
)

The output

----> 1 Choice(
      2     index=0,
      3     finish_reason="stop",
      4     message=msg,
      5 )

File /.venv-3.10/lib/python3.10/site-packages/pydantic/main.py:164, in BaseModel.__init__(__pydantic_self__, **data)
    162 # `__tracebackhide__` tells pytest and some other tools to omit this function from tracebacks
    163 __tracebackhide__ = True
--> 164 __pydantic_self__.__pydantic_validator__.validate_python(data, self_instance=__pydantic_self__)

ValidationError: 1 validation error for Choice
logprobs
  Field required [type=missing, input_value={'index': 0, 'finish_reas...=None, tool_calls=None)}, input_type=dict]
    For further information visit https://errors.pydantic.dev/2.5/v/missing

Setting logprobs to None fixes the problem.

from openai.types.chat.chat_completion import ChatCompletionMessage, Choice

msg = ChatCompletionMessage(role="assistant", content="")

Choice(
    index=0,
    finish_reason="stop",
    message=msg,
    logprobs=None # added line
)

Code snippets

see above

OS

Linux

Python version

Python 3.10.13

Library version

openai 1.6.0

[rattrayalex]

Why are you instantiating Choice in this way? Do you mean to instantiate ChoiceParam?

[rattrayalex]

ah, it seems you're doing this for tests, is that right?

Hmm, we didn't intend for these models to be instantiated this way, so we haven't considered this sort of change to be breaking, but of course in cases like yours, it is.

If we do this we'll want to add = None to all Optional fields.

This is something we'll have to think about and likely won't be actioned until after the new year.

Thank you for raising this!

[davorrunje]

Thank you the reply.

The main reason I raised the issue is that it broke AutoGen (microsoft/autogen#1043). I made a PR there, but until the new version is released everyone has to downgrade openai lib as remedy.

i understand your argument and agree with it in general. The thing that was puzzling me was that the default value was not set it in openai.types.chat.chat_completion.Choice, but it was set in openai.types.chat.chat_completion_chunk.Choice. I could not understand why and I assumed it was an oversight.

[rattrayalex]

Thank you @davorrunje !

The use-case of needing to build up a full object from chunks makes sense. We plan to bake that into this library in the coming weeks, but we should support that.

We do plan to add = None to cases like this in general, so I'll go ahead and merge your PR, as it's a step in the right direction. Thank you!