weiqingy commented on issue #280:
URL: https://github.com/apache/flink-agents/issues/280#issuecomment-5100189302

   
   While working through the Azure step of this issue I looked at whether the 
`azureai` module (Azure AI Inference) can use native structured output. It 
cannot, at the currently pinned `azure-ai-inference:1.0.0-beta.5`, and I want 
to record why so nobody spends the same time on it again.
   
   There are two service-side constraints that contradict each other.
   
   The `json_schema` response format is rejected below api-version 
`2024-08-01-preview`. The service returns `400 "response_format value as 
json_schema is enabled only for api versions 2024-08-01-preview and later"`.
   
   But raising the SDK's service version to `V2024_08_01_PREVIEW` makes the 
endpoint return `404 "Resource not found"`. That 404 is why the SDK pinned its 
own default down: the beta.4 changelog entry is "Use ModelServiceVersion 
(api-version) of 2024-05-01-preview as default". `ModelServiceVersion` has 
exactly two values and `getLatest()` returns the older one.
   
   So one version routes but rejects the feature, and the other accepts the 
feature but does not route. There is no configuration of beta.5 that satisfies 
both.
   
   Worth flagging for anyone who tries: the `2024-05-01-preview` swagger does 
model `ChatCompletionsResponseFormatJsonSchema`, so the code compiles cleanly. 
Compiling is not evidence that the service accepts it.
   
   Azure closed the corresponding SDK issue as won't-fix, pointing at the beta 
SDK's retirement on 2026-08-26 and directing users to the OpenAI v1 API surface 
instead.
   
   I am therefore scoping the Azure work here to the Azure OpenAI connections, 
which use the OpenAI SDK on both the Java and Python sides and support the 
feature normally. Happy to revisit `azureai` if that module ever moves to the 
successor SDK.
   
   
   The Azure OpenAI half is up as #930.
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to