[Feature]: Add max_retries and timeout support to OpenaiChatModel

[ployts]

Description

Currently, the OpenaiChatModel class in OpenJudge does not explicitly handle network timeouts or transient API errors. If the OpenAI API hangs or returns a temporary 5xx error, the OpenaiChatModel might wait indefinitely or crash immediately.

Suggested Implementation

The official OpenAI Python SDK (v1.x) natively supports max_retries and timeout parameters, but OpenJudge is not currently exposing or utilizing them in the model wrapper. This limits the robustness of the evaluation pipeline in unstable network environments.

I would like to expose the native max_retries and timeout parameters in the OpenaiChatModel initialization and pass them to the underlying OpenAI client.

Implementation Steps:

Modify OpenaiChatModel:

Update the init method in openjudge/models/openai_chat_model.py (or the relevant file path) to accept max_retries (int, default e.g., 2) and timeout (float, default e.g., 60.0).

Pass these parameters when initializing the AsyncOpenAI or OpenAI client within the class.

Add Unit Tests:

Create a test case in tests/models/test_openai_model.py.

Use unittest.mock to mock the OpenAI client.

Verify that when OpenaiChatModel is initialized with specific retry/timeout values, those values are correctly passed to the OpenAI client constructor.

Note: The test should not make real API calls; it should only verify parameter passing.

Alternatives

We could implement a custom retry decorator (using libraries like tenacity) around the API call method. However, since the OpenAI SDK v1.x provides built-in support for retries and timeouts, using the native implementation is cleaner and requires fewer dependencies.

Example usage after implementation:

model = OpenaiChatModel(
    model="gpt-4",
    api_key="...",
    timeout=30.0,    # Fail if request takes longer than 30s
    max_retries=3    # Retry 3 times on connection errors
)

[Wangzy455]

Could you assign this to me? Thank you.

[helloml0326]

Sure thing! Assigned to you. Looking forward to your Pull Request! Let us know if you have any questions