AzureChatOpenAI Deployment “o3-mini” Returns 400 Error Due to Unsupported ‘temperature’ Parameter

[keyoumao]

• Package Name: Azure Cognitive Services OpenAI (via AzureChatOpenAI and AzureOpenAIEmbeddings)
• Package Version: langchain-openai = "==0.1.6"
• Operating System: macOS (Apple Silicon, using Homebrew)
• Python Version: 3.10.16

Describe the bug
When using the o3-mini deployment for evaluation, the API returns a 400 error stating:

Unsupported parameter: 'temperature' is not supported with this model.

Despite conditional logic in the code intended to omit the temperature parameter for o3-mini, the parameter is still being sent in some cases. This leads to repeated HTTP POST failures during evaluation.

To Reproduce
Steps to reproduce the behavior:

1. Configure the evaluation to use an o3-mini deployment (ensure that your settings’ model name for either generator or evaluator includes "o3-mini" as a substring).
2. Load a valid checkpoint file when prompted.
3. Run the evaluation command: python evaluate.py
4. Observe the logs where HTTP POST requests to the Azure endpoint return 400 errors with the message regarding the unsupported
   "temperature" parameter.
   Below is a minimal sample code snippet that demonstrates how the client is instantiated. Note that the conditional check should omit the temperature parameter when the model name includes "o3-mini". Despite this, the parameter appears to be sent, triggering the error:

# Sample settings for an o3-mini deployment
settings = {
    "GENERATOR_LLM_ENDPOINT": "https://example.cognitive.microsoft.com/",
    "GENERATOR_LLM_API_KEY": "****",  # Masked API key
    "GENERATOR_LLM_DEPLOYMENT_NAME": "o3-mini",
    "GENERATOR_LLM_MODEL_NAME": "o3-mini",
    "GENERATOR_LLM_API_VERSION": "2025-01-01-preview",
    "GENERATOR_LLM_API_TYPE": "azure",
    "GENERATOR_LLM_TEMPERATURE": 0.7,
}

params = {
    "azure_endpoint": settings["GENERATOR_LLM_ENDPOINT"],
    "deployment_name": settings["GENERATOR_LLM_DEPLOYMENT_NAME"],
    "openai_api_version": settings["GENERATOR_LLM_API_VERSION"],
    "openai_api_key": settings["GENERATOR_LLM_API_KEY"],
    "model_name": settings["GENERATOR_LLM_MODEL_NAME"],
}

# Conditional check to omit 'temperature' for o3-mini
if "o3-mini" not in settings["GENERATOR_LLM_MODEL_NAME"].lower():
    params["temperature"] = settings["GENERATOR_LLM_TEMPERATURE"]

llm = AzureChatOpenAI(**params)
print("LLM instantiated:", llm)

Expected behavior

• The client should not include the temperature parameter when using an o3-mini model deployment.
• The evaluation should proceed without the API returning a 400 error, with a 200 instead.
• Alternatively, if the parameter is necessary, the API should accept it or provide clear documentation on how to handle such requests. We did not have any issues with our current code using gpt-4o or gpt-4o-mini.

Screenshots
. for instance:
2025-03-04 12:52:21,701 - httpx - INFO - HTTP Request: POST https://swedencentral.api.cognitive.microsoft.com/openai/deployments/o3-mini/chat/completions?api-version=2025-01-01-preview "HTTP/1.1 400 model_error" 2025-03-04 12:52:21,703 - ragas.executor - ERROR - Exception raised in Job[2]: BadRequestError(Error code: 400 - {'error': {'message': "Unsupported parameter: 'temperature' is not supported with this model.", 'type': 'invalid_request_error', 'param': 'temperature', 'code': 'unsupported_parameter'}})

Additional context

• The error occurs during one-by-one Q&A evaluation when sending requests to the endpoint.
• This issue blocks the evaluation process and may affect other workflows using the o3-mini deployment.
• Any guidance on either a fix in the API or recommendations for adjusting client requests would be greatly appreciated.

[github-actions]

Thanks for the feedback! We are routing this to the appropriate team for follow-up. cc @trrwilson.

[kristapratico]

Hi @keyoumao can you move this issue to the langchain repo? https://github.com/langchain-ai/langchain/issues

Also, I see you are using an older version of langchain-openai. It looks like this was fixed in 0.2.3 or greater: langchain-ai/langchain@ce33c4f Are you able to upgrade the package?

[github-actions]

Hi @keyoumao. Thank you for opening this issue and giving us the opportunity to assist. To help our team better understand your issue and the details of your scenario please provide a response to the question asked above or the information requested above. This will help us more accurately address your issue.