-
Notifications
You must be signed in to change notification settings - Fork 3.3k
Description
Feature request: Support document-level access control for Azure AI Search tool in Agents SDK
Is your feature request related to a problem? Please describe.
The Azure AI Foundry Agents SDK Azure AI Search tool does not currently support document-level access control (permission trimming) at query time.
Azure AI Search requires the request header x-ms-query-source-authorization (Microsoft Entra token) to enforce document-level security during search queries. While this is supported by the native Azure AI Search SDK, the Agents SDK AI Search tool configuration described here does not expose any way to pass this header or equivalent security context:
https://learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/tools/ai-search
As a result, the built-in AI Search tool cannot be safely used in multi-tenant or per-user access–controlled scenarios.
Describe the solution you'd like
Add support for document-level access control in the Agents SDK Azure AI Search tool, for example by:
- Allowing callers to provide the
x-ms-query-source-authorizationheader - Supporting forwarding of per-request Entra ID / OBO tokens into Azure AI Search
- Providing a documented mechanism for injecting query-time security context
The goal is parity with the underlying Azure AI Search SDK’s document-level security capabilities.
Describe alternatives you've considered
- Calling Azure AI Search directly (outside the Agents SDK) to attach
x-ms-query-source-authorization - Building a custom proxy API in front of Azure AI Search to enforce access control
- Maintaining separate indexes per tenant or user
These approaches add complexity and reduce the value of the built-in AI Search tool.
Additional context
Document-level access control is a core enterprise requirement for RAG and agent-based systems.
Without support for x-ms-query-source-authorization (or an equivalent mechanism), the Azure AI Search tool in the Agents SDK cannot be used safely in production, access-controlled environments.