Clarification about retirement/deprecation of Azure OpenAI On Your Data Extension

[arnabbiswas1]

The latest model that Azure OpenAI's On-Your-Data extension supports is GPT-4o (link). Considering gpt-4o (2024-11-20) is retiring on 2026-06-03,

• There needs to be a declaration if On-Your-Data is also retiring on that date.
• If yes, there needs to be a documented migration path from On-Your-Data to Foundry-Agent-Service.

There are some discussions happening on different repositories (one, two: closed issue). Hence, creating a new issue here.

Open queries about the topic of migration:

• Looks like On-Your-Data works with gpt-4.1, gpt-4.1-mini although the relevant documentation states that these two models are not supported (one, two). In case, gpt-4.1 is declared as a supported model, the retirement date for On-Your-Data shifts to 2026-10-14 (4.1 retirement date) from 2026-06-03 (4o retirement date).
• In the other thread, Foundry Agent Service (new) along with Foundry IQ/Azure AI Search Tool has been recommended as the alternative for On-Your-Data (link). However,

  • The Azure AI Projects client library for Python exposing the Responses API is still in Beta (2025-11-15-preview)
    Reference: link

  • Foundry IQ is also in Beta
    Reference: link

Considering the there is lesser than 6 months of time for the migration and not all the services/APIs are not in GA/Stable, a quick response would be appreciated.

cc: Linda Li (@lindazqli), Sarah Widder (@sarah-widder)

[github-actions]

Thanks for the feedback! We are routing this to the appropriate team for follow-up. cc Travis Wilson (@trrwilson).

[arnabbiswas1]

As of 1st Feb 2026, the retirement date for GPT-4o (2024-11-20) has been shifted to Oct' 2026 (2026-10-01): Retirement Page

As I came to know recently, from the Foundry Discord channel that Foundry Agent Service (new) is still in Public Preview. Considering this fact, postponing this retirement is a welcome relief. Once,

• Foundry Agent Service (new) goes GA
• stable version of Azure AI Projects client library for Python is released
• and Microsoft decides prioritizes between Agentic Retrieval feature in AI Search (Preview), Azure AI Search Tool (from Agent Service: GA?) and FoundryIQ (Preview)

it would be much easier for the developers to decide the migration path from On-Your-Data (Unless microsoft decides it) and then deploy it production.

[arnabbiswas1]

Deviation of Azure AI Search Tool [on top of Foundry Agent Service (New) as of 1st Feb' 2026] compared to Azure Open AI on your data extension:

1. Azure AI Search Tool returns the name of the files which are being used as reference (to generate the response), but not the exact chunks. These chunks need to be supplied as references to the answers (to the end user). Also, these chunks are the only way to evaluate the retrieval process. Reference: 42218
2. For query_type equal to vector, vector_simple_hybrid or vector_semantic_hybrid, a vectorizer needs to be created while creating Azure AI Search Index. This is not documented in Azure AI Search Tool page.
3. For Azure AI Search Tool, filter is defined at the agent level. As per the documentation, filer Applies to all queries the agent makes to the index. As a result, if an application has multiple filters, it needs to maintain multiple agents. New filters would require newer agent instances (This could be a maintenance nightmare since each agent can have multiple version). This is very different from the On-Your-Data, where, filter string can be generated dynamically and can be passed as a string while making chat.completions.create calls.

For 2 and 3 above I will raise two separate issues.

[arnabbiswas1]

Update on 9th Feb' 2026

Azure OpenAI On-Your-Data documentation has been updated with a deprecation notification now.

However, the "Foundry Agent Service" mentioned in the deprecation message points to the older (classic) version of the foundry and not the latest one.

This should be corrected or clarified.

cc: Guy Gregory (@guygregory)

[arnabbiswas1]

Update on 10th Feb' 2026

GPT 4.1 support has been added: