Adding a minimal Azure OpenAI Inference SDK.

[jpalvarezl]

Sorry for not addressing the style checks yet. I am aware of them and will address them ASAP. I just wanted to have this PR opened before the Hackathon was over. Style checks have been addressed.

Contents of this PR

• Minimal definition of ChatCompletions operation. Using as many server default values as possible.
• Support for streamed responses
• Examples using key credentials and AAD
• ChatCompletionsClient can be used with Azure or non-Azure OpenAI service

Remaining TODOs from original PoC

• Would need multipart/form-data support to add audio related operations (happy to contribute this, if that is an option. I have a bad (but working) implementation for this, in the original PoC)
• Add additional fields to ChatCompletions models
• Use code generation tools wherever possible

Edited on 01/10/2024

[heaths]

Please go through style errors. We don't want to increase technical debt, which is why they are build errors. I find it helps to run them locally before pushing more commits, especially since it keeps updating the PR.

On that note, I'm very busy at the moment trying to prepare crates we plan on releasing as betas soon. What timeframe do you have in mind for releasing an OpenAI SDK for Rust? It was not previously on our docket of support. /cc @RickWinter @ronniegeraghty

[heaths]

We're still working on some big changes with azure_core and TypeSpec-branded crates. It sounds as if this crate would have no official support yet - @ronniegeraghty will be contacting you internally - so we'd like to have this in a separate feature branch for now so we're not spending a lot of resources keeping it up to date until its ready for broad support from your team.

I'll do a cursory review about the overall structure and public APIs, but not in-depth at this time. Can you also create a separate feature branch e.g., feature/openai in the Azure/azure-sdk-for-rust repo and retarget your PR (click on the Edit button by the PR title) on that?

git fetch upstream feature/track2 # assumes this is the "upstream" repo
git push upstream FETCH_HEAD:feature/openai
git checkout jpalvarezl/aoai_from_track2 # check out your topic branch if not already
git rebase upstream/feature/openai
git push --force

Then in your PR on this site, retarget using the Edit button.

[heaths]

You should also implement azure_core::ClientOptionsBuilder. See guidelines for details. We'll have a derive macro eventually to make this easier but aren't there yet.

[heaths]

Same is true for client method options.

[heaths]

As a caller, I would wonder what the difference is between endpoint and base_url. Public API docs will be required before you can ship, but I would describe at least these two methods now.

[heaths]

These should be client method options. See guidelines.

[heaths]

This is clearly going through a lot of development and iteration. I highly recommend building and testing locally before submitted. As it's not ready for review (see comments above, along with how frequently you're pushing commits) I'm marking this as a draft. When you think it's ready for a final review and have talked with @ronniegeraghty, please convert it back to a full PR.

[jpalvarezl]

This is clearly going through a lot of development and iteration. I highly recommend building and testing locally before submitted. As it's not ready for review (see comments above, along with how frequently you're pushing commits) I'm marking this as a draft. When you think it's ready for a final review and have talked with @ronniegeraghty, please convert it back to a full PR.

My apologies if it seemed like I was adding more changes to this PR. I tried following the steps you suggested to move the branch to this repository, but I must have done something wrong and ran into some issues which resulted on me ending up doing something else entirely by accident, causing a mess with the commit history. It seems like I rebased Azure/feature/track2, which was injected as a commit at what was HEAD at the time of trying. I will sort this out.

First of all, thank you for your early feedback. I have addressed all the items that I marked as resolve taking in your comments in full. Also, addressed the CI checks. The comments that are still unresolved, seemed a bit more involved, so I want to find the time to be able to dedicate my full attention to them.

I am waiting on being able to talk sync with @trrwilson to figure out how to best approach and move forward with this effort.