Spike: Test Azure OpenAI via OpenAI different base URL + API key

[jthomasmock]

Goal: Confirm Azure OpenAI support via OpenAI with alternate base URL when using an API key.

• Can we use the existing handling with provider openai-api or Custom Provider openai-compatible to support this? If not, what errors do we get and what changes might we need?

[samclark2015]

I was able to start preliminary testing of this. At face value, it seems likely this will work with the existing OpenAI-compatible provider with no/minimal changes.

After creating an Azure AI Foundry project, an OpenAI compatible endpoint URL + API key was provided. Adding this to Positron resulted in a successful call to fetch models (via the /openai/v1/models endpoint); however, all models that Azure provides was returned. This is in conflict with their workflow of "deploying" a model, and so no models were actually able to be used without being deployed.

After deploying a model (gpt-5, gpt-oss-120b), the call to the /models endpoint now fails with the following error:

Access denied due to invalid subscription key or wrong API endpoint. Make sure to provide a valid key for an active subscription and use a correct regional API endpoint for your resource

I'll explore further to determine the cause of this error and a possible solution.

[samclark2015]

Follow-up

The above error seems to have been a red herring; it was resolved after a few attempts at configuring Azure as a provider.

Analysis

Azure's OpenAI API does not have a mechanism to return only deployed models, so will not work as an OpenAI-compatible provider without additional work.

PR #10611 allows users to manually configure models - avoiding a call to the problematic /models endpoint. When deployed models are configured using that PR, Azure works as expected!

To configure in Positron Assistant:

1. Configure positron.assistant.configuredModels to enumerate models deployed on Azure under the openai-compatible provider.
2. An OpenAI-compatible endpoint must be configured as https://[Azure OpenAI endpoint]/openai/v1 - where Azure OpenAI endpoint is the URL from the AI Foundry portal.

Future work

• Implement dedicated Azure model provider
• Add logic to properly query deployed models from Azure endpoints