Azure Open AI Async filter results in None type content part

[martgra]

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

When using the async filter in Azure, stream mode, the SDK returns None types. Unsure whether this is a issue in the API or the SDK but since this is a fundament package for many high level libraries the fix could be implemented here.

Ref other reported issues:
pydantic/pydantic-ai#797
microsoft/semantic-kernel#7250
langchain-ai/langchain#25951

To Reproduce

1. Set Azure Open AI completion model (GPT-4x) to Async content filter (stream mode)
2. Run stream with async client
3. Try accessing content gives unexpected None type.

Error will only occur with async content filter.

Ref error:

  File "/workspace/backend/api/services/domain_service.py", line 50, in get_answer
    async for response in self.message_service.get_grounded_message_as_stream(
  File "/workspace/backend/api/services/message_service.py", line 83, in get_grounded_message_as_stream
    async for chunk in self.chat_consumer.get_message_as_stream(message_history_with_context):
  File "/workspace/backend/api/consumers/chats/azure_openai_consumer.py", line 56, in get_message_as_stream
    content = chunk.choices[0].delta.content
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'NoneType' object has no attribute 'content'

Code snippets

async with await client.chat.completions.create(
                messages=messages,
                model="gpt-4o",
                stream=True,
            ) as stream:
                async for chunk in stream:  
                    if not (
                        chunk
                        and chunk.choices
                    ):
                        continue

                    content = chunk.choices[0].delta.content

OS

Debian Bookworm, Mac OS

Python version

Python v3.12

Library version

open ai v.1.69.0

[kapis]

I believe this is a problem on the Azure side - for some reason, they've decided to violate the OpenAI spec with the async filter.

See this: #2255