Openai service got an unknown error

Hello noppavut,

Greetings! Thanks for raising this question in the Q&A forum.

This is a generic server_error (HTTP 500) from the Azure OpenAI backend, not something wrong in your code. It means the service failed while generating a response rather than rejecting your request for a configuration reason, so it is not something you can fix by changing your payload. It's also worth noting that GPT-5.2 and Bing Search grounding through the Responses API are both fairly new, and 500-class errors like this are currently being reported by multiple customers combining recent GPT-5.x models with the Responses API and tool calling, most often tied to transient backend capacity pressure rather than a lasting outage.

Here is how to work through it:

1. Retry with exponential backoff. The official SDKs already retry connection errors, 408, 429, and >=500 errors twice by default, but that's often not enough during a capacity blip. Increase retries explicitly:

from openai import AzureOpenAI

client = AzureOpenAI(
    azure_endpoint=azure_endpoint,
    api_key=azure_key,
    api_version="2025-04-01-preview",
    max_retries=5,
    timeout=60.0,
)

This is a generic server_error (HTTP 500) from the Azure OpenAI backend, not something wrong in your code. It means the service failed while generating a response rather than rejecting your request for a configuration reason, so it is not something you can fix by changing your payload. It's also worth noting that GPT-5.2 and Bing Search grounding through the Responses API are both fairly new, and 500-class errors like this are currently being reported by multiple customers combining recent GPT-5.x models with the Responses API and tool calling, most often tied to transient backend capacity pressure rather than a lasting outage.

Here is how to work through it:

1. Retry with exponential backoff. The official SDKs already retry connection errors, 408, 429, and >=500 errors twice by default, but that's often not enough during a capacity blip. Increase retries explicitly:

from openai import AzureOpenAI

client = AzureOpenAI(
    azure_endpoint=azure_endpoint,
    api_key=azure_key,
    api_version="2025-04-01-preview",
    max_retries=5,
    timeout=60.0,
)

If you're on the streaming Responses API, also handle the case where a 200 OK is returned and streaming starts normally, but a mid-stream error event is emitted partway through generation. This is expected behavior for transient capacity issues and needs to be caught and retried at the application layer, not just at the HTTP level.

Isolate whether the Bing grounding tool is the trigger. Send the same prompt through the Responses API with the tools parameter removed (no Bing grounding). If the plain call succeeds consistently while the tool-enabled call fails, this points to a tool/model compatibility issue rather than general capacity, which is worth flagging explicitly in your support request.

Check Azure Service Health for your region. Generic 500s with a request ID are frequently tied to regional capacity incidents that don't always show up on the public Azure status page but do show up under Service Health > Resource Health in the portal for your specific resource. Confirm which region your resource is deployed in and check there first.

If possible, test the same deployment in another supported region (for example East US 2, Sweden Central, or wherever GPT-5.2 is available in your subscription). If the identical request succeeds elsewhere, that confirms a region-specific capacity issue rather than something wrong with your integration.

If it persists across retries, tools removed, and regions, this needs backend investigation from the product team. Open an Azure support request and include the request ID from the error message (af74237a-74d3-4f87-8ce3-5996043571fb), your deployment name, region, and whether removing the Bing grounding tool changes the outcome. That combination of details lets Support trace the exact backend failure quickly.

If this answer helps you kindly accept the answer which will help others who have similar questions.

Best Regards,

Jerald Felix.