ai icon indicating copy to clipboard operation
ai copied to clipboard
verified publisher icon vercel

📯 Azure community team updates

Open gr2m opened this issue 1 month ago • 16 comments

Hi there 👋🏼

This is a follow up to https://github.com/vercel/ai/pull/8238#issuecomment-3339970839.

We tried to setup @vercel/ai-sdk-azure with outside collaborators, but unfortunately that is not working. So instead we can use this issue to collaborate on all things @ai-sdk/azure.

Our Azure Community Team is

user timezone
@gr2m UTC-8
@jephal UTC+1
@rahulbhadja UTC-5
@tommasoghisini UTC+1
@tsuzaki430 UTC+9

Relevant issues and pull requests are labeled with https://github.com/vercel/ai/labels/provider%2Fazure

Monthly meeting: every 2nd Thursday at 16:30 UTC Next call: December 11th at 16:30 UTC / 8:30 PST / 17:30 CET / 11:30 ET (zoom link)

gr2m avatar Oct 28 '25 19:10 gr2m

Question: should we make azure.responses the default instead of azure.chat? 💬 https://github.com/vercel/ai/issues/8596#issuecomment-3458458336 👀

gr2m avatar Oct 28 '25 21:10 gr2m

Thanks @gr2m for creating a way to communicate! Yes, let me start working on this issue. I also created one PR for an issue related to telemetry, and another PR for the AssemblyAI provider. I’m waiting for your review!

rahulbhadja avatar Oct 28 '25 21:10 rahulbhadja

I've been experimenting a bit with model router from azure and created #9966. Not somthing that needs to be urgently addressed but more documenting a few things about it (in case you dont want to keep issue open i can convert it to discussion).

https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/model-router

tommasoghisini avatar Nov 01 '25 18:11 tommasoghisini

@jephal @rahulbhadja @tommasoghisini @tsuzaki430

  1. As of today, we have 73 open issues/PRs for Azure: https://github.com/vercel/ai/labels/provider%2Fazure

    I'd love get to 0 by the end of this year (except the ones that are not actionable, if any).

    When you have a chance, could you skim through the list and share what you think we should prioritize?

  2. If you are comfortable sharing: what timezone are you in? I'd like to start an "AI SDK + Azure" office hour, where I'd work on Azure issues and invite everyone to join to help out and socialize. I'd love to find the time that is convenient for everyone

gr2m avatar Nov 05 '25 19:11 gr2m

Hi @gr2m, sure, I'm in the EST time zone. I'm pretty flexible with office hours and usually work most of the time, haha. Regarding issues related to Azure, most of them are old and already fixed, but I will start going through them again.

rahulbhadja avatar Nov 05 '25 20:11 rahulbhadja

Heyo @gr2m sounds good! I'm in UTC +1.

Anything after 4pm my time, i should be able to join.

Ill also take a look at the list of issues/prs and try to see if i find something to prioritze or that we might be able to close

jephal avatar Nov 05 '25 21:11 jephal

Hello😊 Thank you messages.

UTC+9 , in Japan. However, I apologize. I am not very good at English, and I am afraid that attending the office hours would be a nuisance to everyone, so I would like to request not to participate. I am truly sorry. I would like to check the provider/azure issur/pr list. Thank you very much for your understanding.

tsuzaki430 avatar Nov 06 '25 00:11 tsuzaki430

Of course, no problem at all @tsuzaki430. You can still share your priorities in writing ahead of office hours, and we will look into them.

gr2m avatar Nov 06 '25 15:11 gr2m

@jephal @rahulbhadja @tommasoghisini

time wise: would 17:00–18:00 UTC work? (12:00–13:00 @rahulbhadja, 18:00–19:00 @jephal/@tommasoghisini). I could also start 30-60min earlier, it would just make it a bit harder with my family logistics, but happy to do it if otherwise one of you couldn't attend.

Would you prefer Wednesday or Thursday?

I'd suggest a monthly check in starting next week

gr2m avatar Nov 06 '25 16:11 gr2m

12.00-1:00 PM works for me. I'm available on both Wednesday and Thursday, thanks @gr2m

rahulbhadja avatar Nov 06 '25 17:11 rahulbhadja

Would prefer thursday, but can also do wednesday.

If we could do 1 hr before that would be optimal, but I can make do with 18-19 :)

Thank you @gr2m !

jephal avatar Nov 07 '25 10:11 jephal

same here :)

tommasoghisini avatar Nov 07 '25 11:11 tommasoghisini

Okay great, let's meet every 2nd Thursday at 16:30-17:30 UTC. I'll update the issue description with a link

I went through all Open Azure issues and pull requests. They are mostly older and probably no longer relevant, but it would be good to go through and verify if the problems are still replicable.

In summary, I feel there like people are asking for better observability and a way to pass a model ID when a deploymend ID is passed to the ai methods, e.g. see https://github.com/vercel/ai/issues/7383

gr2m avatar Nov 07 '25 23:11 gr2m

Yes, I saw many issues regarding observability and telemetry not passing the correct modelId. I also tried to solve the modelName and deploymentName issue #9497, but yes, definitely we can discuss more regarding this as well.

rahulbhadja avatar Nov 08 '25 04:11 rahulbhadja

Community Meeting from November 13, 2025

Notes from yesterdays call with @jephal, @tommasoghisini, and @rahulbhadja

Intros

  • Jeppe and Tommaso are coworkers at falck.com. They use the AI SDK and the Azure Provider for an internal chat app.
  • Rahul is Co-founder of a startup using AI SDK with Azure-hosted models

Learnings

  • Azure is significantly different from other providers, making it complex to set up and debug
  • Azure frequently changes naming conventions and repackages their AI services
  • Content filtering and API versioning present additional challenges

Priorities

  • Code Interpreter File Downloads: Users can't easily download files generated by the Code Interpreter tool. We have an open pull request by @tsuzaki430 that needs to be updated: https://github.com/vercel/ai/pull/9135
  • Telemetry Issues:
    • short term: fix discrepancy between deployment ID and model ID in observability logs
    • longer term: observability will be a priority for the AI SDK soon, we will incorporate requirements by Azure
  • Documentation is out of date since azure.responses was made the default: https://github.com/vercel/ai/pull/10209

Next

  • Azure Deep Research Integration: @tommasoghisini is exploring Azure's deep research setup with GPT-5 and grounding with Bing and will document findings to gauge community interest
  • Interest in real-time models for live interactions, but these are not yet prioritized by AI SDK

gr2m avatar Nov 14 '25 17:11 gr2m

Hello!Thank you for sharing the meeting results.

Intros

Thank you for telling me about everyone's activities.Learning about them made me feel energized.I am engaged in a small project developing web applications that can utilize Azure both internally and externally.

Learnings

Regarding the difficulty of Azure, I felt the same way.Error handling is something I personally find challenging. I once built an error UI for content filters, but it no longer works.I'm not sure of the cause, but it might be related to the API version.Regardless, I think Azure is a difficult provider.

Priorities

  • Understood regarding the Code Interpreter. Thank you for waiting.I have implemented the basic functionalities of Code Interpreter in #9135, but I had many concerns about the impact of the openai package and whether the changes, made solely for my convenience, would negatively affect others.However, since most of those major concerns have been resolved in main by other PRs, I think I can move forward.I might break it down into smaller PRs and recreate it.

  • I understand that the telemetry issue takes priority. I am very grateful that everyone is tackling the problem where the relationship between deployment ID and model ID is causing various negative impacts.

Next

  • Since the Bing Search v7 API ended in August of this year, I had to discontinue the web search tool I developed myself.Azure Deep Research Integration would be very useful for me, and I believe others would be very interested as well.
  • The real-time API is a very futuristic feature, and I'm very excited about its future.Currently, text-based communication remains important, so I understand the prioritization.

tsuzaki430 avatar Nov 15 '25 01:11 tsuzaki430

Hello , I created issue #10255 for CodeInterpreter output files. We can check the latest status for it😊

tsuzaki430 avatar Nov 15 '25 13:11 tsuzaki430

Hey guys, Azure added Anthropic models in Foundry source.

Do we have any plans to support this new addition? I checked the SDK structure, and it follows the same pattern as the Anthropic native SDK. @gr2m

rahulbhadja avatar Nov 19 '25 05:11 rahulbhadja

I assume it could work similar to how we currently implement anthropic in the amazon-bedrock and google-vertex provider

gr2m avatar Nov 19 '25 07:11 gr2m

Hello,

There have been several updates to the built-in tools in the Responses API. I have created new pull requests for the following work🚀

tsuzaki430 avatar Nov 20 '25 06:11 tsuzaki430

The example for Anthropic Claude APIs on Microsoft Foundry using@ai-sdk/anthropic😊

#10400

tsuzaki430 avatar Nov 20 '25 13:11 tsuzaki430

You're killing it @tsuzaki430, thank you!!

gr2m avatar Nov 20 '25 17:11 gr2m

I’ve created two PRs related to CodeInterpreter.

  • #10252

  • #10266

If you have any feedback regarding the broader impact of these two PRs, please leave your comments in issue #10255. I’d appreciate it if you could take a look when you have time.

tsuzaki430 avatar Nov 22 '25 14:11 tsuzaki430

Could someone else also review @tsuzaki430's latest pull request?

gr2m avatar Nov 23 '25 20:11 gr2m

sorry been busy at work, will review it now @gr2m @tsuzaki430 #10266

tommasoghisini avatar Dec 04 '25 10:12 tommasoghisini