Python: Bug: Structured Output fails when ChatCompletionAgent is configured with AzureAIInferenceChatCompletion

[evmin]

Describe the bug

When following the instructions [1] everything operational when run AS-IS. However if I switch to use AzureAIInferenceChatCompletion, the code fails with JSONDecodeError. The model does provide a correct response, but the invocation does not return Structured Output.

Attaching the code sample to reproduce

[1] https://github.com/microsoft/semantic-kernel/blob/main/python/samples/getting_started_with_agents/chat_completion/step10_chat_completion_agent_structured_outputs.py

To Reproduce Steps to reproduce the behavior:

1. Open the referenced file [1]
2. Update the file to use "AzureAIInferenceChatCompletion" (see below the code sample)
3. Run the file an see the error

Expected behavior To execute and produce the output as per example in the file - which is a copy of [1]

Platform

• Language: Python
• Source: [e.g. NuGet package version 0.1.0, pip package version 0.1.0, main branch of repository]
• AI model: Azure OpenAI: GPT-4.1
• IDE: VS Code
• OS: Mac

AzureAIInferenceChatCompletion

    # ----------------------------------------------------------------------
    # Commented out: when using AzureChatCompletion, code works.
    # ----------------------------------------------------------------------
    # service=AzureChatCompletion(
    #     endpoint=endpoint,
    #     api_version=api_version,
    #     api_key=api_key,
    #     deployment_name=MODEL_NAME,
    # )
    
    # ----------------------------------------------------------------------
    # When using AzureAIInferenceChatCompletion, code fails
    # ----------------------------------------------------------------------
    from semantic_kernel.connectors.ai.azure_ai_inference import AzureAIInferenceChatCompletion
    from azure.ai.inference.aio import ChatCompletionsClient
    from azure.identity.aio import DefaultAzureCredential
       
    service = AzureAIInferenceChatCompletion(
        ai_model_id="executor",
        service_id="executor",
        client=ChatCompletionsClient(
            endpoint=f"{str(api_base).strip('/')}/openai/deployments/{MODEL_NAME}",
            api_version=api_version,
            credential=DefaultAzureCredential(),
            credential_scopes=["https://cognitiveservices.azure.com/.default"]
        ))
    ```

[moonbox3]

The AzureAIInference SDK only supports "json_object" mode and not structured outputs. https://github.com/microsoft/semantic-kernel/issues/11025#issuecomment-2789106777

[TaoChenOSU]

Closing in favor of: https://github.com/microsoft/semantic-kernel/issues/11952