Python: Bug: AttributeError in azure_ai_agent when invoking agent with RAG (Azure AI Search) on v1.32.2

[MarekFabinski]

Description
After updating to the latest version of the semantic-kernel library, invoking an AzureAIAgent that is configured to use the built-in Azure AI Search tool for Retrieval-Augmented Generation (RAG) consistently fails. The process throws an AttributeError: 'RunStepDeltaToolCall' object has no attribute 'azure_ai_search' deep within the library's streaming logic.

This appears to be a breaking change or a bug in how the latest SDK version parses streaming tool call responses from the Azure AI service, as downgrading to an earlier version 1.32.1 resolves the issue.

Expected behavior
The invoke() call should complete successfully. The AzureAIAgent should use its RAG tool in the cloud, synthesize an answer, and stream the final ChatMessageContent back to the orchestrator without any errors.

[moonbox3]

Thanks for filing, @MarekFabinski. Must be related to the Agent service's move to GA and not captured within SK. Will get this handled ASAP.

[sumansuhag]

Hey MarekFabinski,

Thanks for reporting that important bug! It looks like you ran into an AttributeError that popped up after the update from version 1.32.1 to 1.32.2 in the semantic-kernel library, which is definitely a problem when dealing with RAG and Azure AI Search.

Your guess about it being a breaking change or a bug in how the latest SDK version parses streaming tool call responses sounds spot on. The RunStepDeltaToolCall objects are part of the output from agents, and if there’s been a change in how their structure is expected or how the SDK reads them, that could cause this kind of error.

The fact that going back to 1.32.1 fixes the issue really points to a problem that came with 1.32.2.

This needs to be looked at quickly by the semantic-kernel team. They should compare the changes between these versions, especially anything related to agent streaming and tool call handling, as well as the azure_ai_search integration. It could be a regression or a breaking change that didn’t get documented well.

For now, it’s best to stick with 1.32.1. I’m looking forward to seeing the official fix, as having RAG work smoothly with Azure AI Search is really important for a lot of projects.

[seizeuniverse]

Hello, when this issue will be handled, should we expect it to be included in 1.35.0 @moonbox3 ?

[moonbox3]

@seizeuniverse PR open with the fix.