openai[patch]: enable Azure structured output, parallel_tool_calls=Fa…

[baskaryan]

…lse, tool_choice=required

response_format=json_schema, tool_choice=required, parallel_tool_calls are all supported for gpt-4o on azure.

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

1 Skipped Deployment

Name	Status	Preview	Comments	Updated (UTC)
langchain	⬜️ Ignored (Inspect)	Visit Preview		Sep 18, 2024 0:48am

[Skar0]

Hi @baskaryan, thank you for addressing this issue and for your work on this! I tested the changes, and they work as expected for the json_schema part 😃

I noticed that in your fix, AzureChatOpenAI._create_chat_result now passes an openai.BaseModel to BaseChatOpenAI._create_chat_result instead of the dict resulting from response.model_dump(), as it did before. While this does resolve the issue for which I proposed a fix in #26086, it requires both BaseChatOpenAI._create_chat_result and AzureChatOpenAI._create_chat_result to transform the response from an openai.BaseModel into a dict.

In #26086, I proposed modifying BaseChatOpenAI._create_chat_result to consistently work with the dict version of the response, which is always created anyways. This change eliminates the need to convert the response into a dict multiple times when calling AzureChatOpenAI._create_chat_result and I believe it simplifies the logic and complexity within these methods. I believe that this would complement your changes and improve code clarity. Of course, this is assuming that my proposed changes are correct and align with your goals for the implementation of that method.

Thanks again for your work and for considering my contribution! 😉

[ccurme]

Ran OpenAI integration tests on this branch and got failures related to env vars: https://github.com/langchain-ai/langchain/actions/runs/10924401318/job/30323353395

Also getting some failures here in json mode on master, suspect unrelated. Unable to reproduce so far.

FAILED tests/integration_tests/chat_models/test_azure.py::test_json_mode_async - openai.LengthFinishReasonError: Could not parse response content as the length limit was reached - CompletionUsage(completion_tokens=10, prompt_tokens=18, total_tokens=28, completion_tokens_details=None)

[baskaryan]

huh thought id set that env var, will double check

[baskaryan]

Hi @baskaryan, thank you for addressing this issue and for your work on this! I tested the changes, and they work as expected for the json_schema part 😃

I noticed that in your fix, AzureChatOpenAI._create_chat_result now passes an openai.BaseModel to BaseChatOpenAI._create_chat_result instead of the dict resulting from response.model_dump(), as it did before. While this does resolve the issue for which I proposed a fix in #26086, it requires both BaseChatOpenAI._create_chat_result and AzureChatOpenAI._create_chat_result to transform the response from an openai.BaseModel into a dict.

In #26086, I proposed modifying BaseChatOpenAI._create_chat_result to consistently work with the dict version of the response, which is always created anyways. This change eliminates the need to convert the response into a dict multiple times when calling AzureChatOpenAI._create_chat_result and I believe it simplifies the logic and complexity within these methods. I believe that this would complement your changes and improve code clarity. Of course, this is assuming that my proposed changes are correct and align with your goals for the implementation of that method.

Thanks again for your work and for considering my contribution! 😉

the issue is that we want to extract response.choices[0].message.parsed before response is converted to a dict, as parsed can itself be a pydantic object and we don't want it converted to a dict