[WIP -DO NOT REVIEW] AI: Adds AzureOpenAIEmbeddingGenerator provider for issue #5842

[ananth7592]

Summary

Implements AzureOpenAIEmbeddingGenerator — the concrete Azure OpenAI-backed implementation of ICosmosEmbeddingGenerator that lives in the Microsoft.Azure.Cosmos.AI package.

Closes #5842.

Dependencies

This PR cherry-picks from three unmerged dependency PRs. It will be rebased / squashed onto main once those land:

• [Internal] VectorEmbeddingPolicy: Adds EmbeddingSource block to Embedding model #5848 — EmbeddingSource / EmbeddingAuthType model (Arooshi)
• EmbeddingGenerator: Adds ICosmosEmbeddingGenerator client-wide configuration (preview) #5838 — ICosmosEmbeddingGenerator interface on CosmosClientOptions / QueryRequestOptions
• Embedding: Adds Microsoft.Azure.Cosmos.AI package scaffold #5845 — Microsoft.Azure.Cosmos.AI package scaffold

Changes

Microsoft.Azure.Cosmos.AI/src/AzureOpenAIEmbeddingGenerator.cs

New sealed class implementing ICosmosEmbeddingGenerator and IDisposable.

Constructors (4 public)

• (EmbeddingSource source, int dimensions, string apiKey) — builds from the VectorEmbeddingPolicy's source block + API key
• (EmbeddingSource source, int dimensions, TokenCredential credential) — same but Entra auth
• (string endpoint, string deploymentName, int dimensions, string apiKey) — explicit params + API key
• (string endpoint, string deploymentName, int dimensions, TokenCredential credential) — explicit params + Entra auth

Behavior

• Batches inputs at the Azure OpenAI hard limit of 2048 per call; results concatenated in original order
• Creates a fresh EmbeddingGenerationOptions { Dimensions = ... } per batch call (the Azure SDK may mutate options internally during concurrent serialization)
• Wraps RequestFailedException → CosmosException(503) with endpoint/deployment in the message
• Validates inputs up-front: null collection, null elements, empty or whitespace-only elements each throw ArgumentException
• Thread-safe: the underlying EmbeddingClient is stateless for concurrent use

Testability

An internal test constructor accepts a Func<IReadOnlyList, int, CancellationToken, Task<IReadOnlyList<ReadOnlyMemory>>> delegate, bypassing the sealed EmbeddingClient entirely. This is consistent with the approach used elsewhere in the SDK for injecting fake network calls.

Microsoft.Azure.Cosmos.AI/tests/.../AzureOpenAIEmbeddingGeneratorTests.cs

25 unit tests covering:

• Constructor validation (null/empty/whitespace endpoint, deployment, key; zero/negative dimensions; null/invalid EmbeddingSource)
• GenerateEmbeddingsAsync input validation (null, empty collection, null elements, whitespace/empty elements)
• Happy path (single batch, correct values, correct order)
• Dimensions forwarded correctly to the batch func
• CancellationToken forwarded correctly
• Batching: 5000 inputs → 3 calls (2048 + 2048 + 904); exact 2048 → 1 call
• Multi-batch order preservation (2050 inputs, numeric marker in each vector)
• Cancellation propagation (TaskCanceledException)
• Exception propagation from the batch func
• ObjectDisposedException after Dispose()
• Double-dispose is safe

[ghost]

All good!

[Pilchie]

In the .NET org, they have a task to generate this at build time so that you don't have to look at the PublicKey in source, not important, as this works too, but might be interesting to look at.

[Pilchie]

.NET 6 is out of support. We should target at least .NET 8 here.

[Pilchie]

Should these just be built as part of the regular SDK pipelines? What the scenario for having separate pipelines?

[Pilchie]

We don't appear to actually anything useful in Dispose, and IDisposable types are a pain. Let's not make it IDisposable unless we know of a reason to.

[Pilchie]

Suggested change

	IEnumerable<string> text,
	IEnumerable<string> texts,

nit: It's common for enumerables to be pluralized.

[Pilchie]

In theory we could parrallelize all of these by not awaiting here, and instead collect the tasks into a list/array and do a await Task.WhenAll(tasks) on the results.

Can be done later if we want.

[Pilchie]

Suggested change

	/// <summary>
	/// Releases resources held by this instance.
	/// </summary>
	public void Dispose()
	{
	this.disposed = true;
	}

[Pilchie]

In theory you can get rid of this if and just do the loop below, which will only run once.

[Pilchie]

Is doing this loop useful over just getting the wrapped exception from the service?

[Pilchie]

Suggested change

	private void ThrowIfDisposed()
	{
	if (this.disposed)
	{
	throw new ObjectDisposedException(nameof(AzureOpenAIEmbeddingGenerator));
	}
	}

[ananth7592]

Kevin, Thanks for your review. While all the files you reviewed in this PR are still relevant, this PR is still in a WIP state, I should have marked it in WIP. Will keep this posted and reply back once its ready for review .