AI Foundry started erroring connecting to AI Search

The symptoms indicate an authorization/connection issue between Foundry Agent Service and Azure AI Search that started recently and affects all agents using AI Search.

Based on the documented troubleshooting patterns, the following checks and mitigations apply:

1. Validate AI Search connection authentication
   • If using keyless (Microsoft Entra managed identity) authentication from Foundry to Azure AI Search, confirm that the Foundry project’s managed identity still has the required roles on the search service:
     • Search Index Data Contributor
     • Search Service Contributor
   • If using key-based authentication, confirm that the API key configured in the project connection is still valid and enabled on the Azure AI Search resource.
2. Recreate the project connection to Azure AI Search
   • A broken or stale connection definition can cause repeated token/ARA failures.
   • Recreate the project connection to the Azure AI Search resource using the supported mechanism for Foundry projects (not the legacy ML workspace tooling):
     • Use az cognitiveservices account project connection create or the azure-mgmt-cognitiveservices SDK, ensuring the endpoint and index name are correct and point to the intended search service and index.
3. Confirm endpoint, index, and network configuration
   • Verify the connection configuration:
     • Endpoint is the correct Azure AI Search service URL that actually hosts the index.
     • Index name exactly matches the index in Azure AI Search (case-sensitive).
   • If the Azure AI Search resource is on a private virtual network and the connection uses key-based authentication, switch the connection to Microsoft Entra project managed identity (keyless). Key-based auth is not supported with private networking for agents and can surface as connection/ARA failures.
4. Check for authorization failures from the Foundry side
   • For 401/403 or related authorization issues when Foundry connects to a knowledge base backed by Azure AI Search:
     • Ensure the project’s managed identity has at least Search Index Data Reader on the search service (and Data Contributor if writing to indexes).
     • Ensure the user or service principal creating or updating the project connection has permissions on the Microsoft Foundry resource and project.
     • Confirm the environment is signed in to the correct tenant and subscription when using keyless auth.
5. Validate that the agent is correctly configured to ground on the knowledge base
   • After fixing the connection, confirm the agent configuration still includes the knowledge base tool and that instructions require using it for grounding. If the agent temporarily works after removing/re-adding the knowledge base, but fails again after reload, this strongly suggests the persisted connection object is invalid or missing permissions; recreating the connection at the project level and reattaching it to the agent is required.
6. For the inability to create new agents
   • The same underlying Foundry resource or project-level permission/connection issue can block both new agent creation and AI Search access.
   • Verify that the principal creating agents has the necessary permissions on the Microsoft Foundry resource and project, and that there are no outstanding authorization failures from Azure Resource Manager when creating or deleting project connections.

If, after verifying roles, keys, endpoints, and recreating the AI Search project connection, the ARA BadRequest persists across multiple agents and environments, capture the failing request ID and open a support ticket so the platform team can inspect the ARA/Foundry backend logs.

References:

• Connect an Azure AI Search index to Foundry agents (python)
• Connect an Azure AI Search index to Foundry agents (csharp)
• Connect an Azure AI Search index to Foundry agents (rest)
• Connect an Azure AI Search index to Foundry agents (typescript)
• Connect a Foundry IQ knowledge base to Foundry Agent Service
• Troubleshooting common indexer errors and warnings in Azure AI Search