fix(azure): forward api_version to aembedding() for Azure AI Foundry v1 endpoints

[Sameerlite]

What

Fixes a silent parameter drop in the async embedding path for Azure AI Foundry.

Root Cause

In BaseAzureLLM.embedding(), when aembedding=True, the call to self.aembedding() was missing api_version=api_version. This meant get_azure_openai_client() received None instead of "v1", causing _is_azure_v1_api_version() to return False and AsyncAzureOpenAI to be selected instead of AsyncOpenAI. The wrong client constructs Azure-specific URLs that don't exist on AI Foundry endpoints, resulting in a 404 ResourceNotFound.

The sync path was unaffected — it passed api_version directly to get_azure_openai_client(). Only the async path (which the proxy always uses) had this bug.

Fix

One-line fix in litellm/llms/azure/azure.py: add api_version=api_version to the self.aembedding() call.

Tests

Added tests/litellm/llms/azure/test_azure_embedding.py with:

• Regression test verifying api_version is forwarded through embedding() → aembedding()
• Tests verifying get_azure_openai_client() returns AsyncOpenAI (not AsyncAzureOpenAI) for api_version="v1"
• Tests verifying the /openai/v1/ base URL is used for v1 clients
• Parametrized test covering all v1 variants ("v1", "latest", "preview")

Fixes #24848

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Apr 1, 2026 10:57am

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing Sameerlite:litellm_azure-v1-async-embed-fix (f9e643a) with main (33c3f13)}

[greptile-apps]

Greptile Summary

This PR fixes a silent parameter drop in the async embedding path for Azure AI Foundry: api_version was not being forwarded from embedding() to aembedding(), causing get_azure_openai_client() to receive None and select AsyncAzureOpenAI instead of AsyncOpenAI for v1 endpoints, resulting in 404 errors on AI Foundry. The fix is a single-line addition of api_version=api_version to the self.aembedding() call.

• Core fix (litellm/llms/azure/azure.py): Adds the missing api_version=api_version kwarg to the self.aembedding() dispatch, ensuring _is_azure_v1_api_version() returns the correct value and the right client type is chosen for the async path.
• Tests (tests/litellm/llms/azure/test_azure_embedding.py): Four mock-only unit tests covering the regression directly, correct client-type selection, base URL format, and all v1 api_version variants (\"v1\", \"latest\", \"preview\"). No real network calls are made, satisfying the CI test policy.
• Minor: The sys.path.insert in the new test file uses 5 .. segments (landing one level above the repo root) instead of the correct 4; tests still pass in practice due to pytest's rootdir detection, but the path is technically wrong.

Confidence Score: 5/5

Safe to merge — the fix is minimal, correct, and well-tested with mock-only unit tests.

Only a single P2 finding (incorrect sys.path depth in the test file) remains, which does not affect test execution or production behavior. The bug fix itself is a straightforward one-liner with proper regression coverage.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/azure/azure.py	One-line fix: adds api_version=api_version to the self.aembedding() call so the async embedding path selects the correct OpenAI client for Azure AI Foundry v1 endpoints. Change is correct and minimal.
tests/litellm/llms/azure/test_azure_embedding.py	New mock-only test file with 4 tests covering the regression, client type selection, base URL, and all v1 api_version variants. Minor issue: sys.path insert uses 5 .. instead of the correct 4.
tests/litellm/llms/azure/init.py	Empty package init file, no issues.

Sequence Diagram

sequenceDiagram
    participant Caller
    participant Embed as embedding()
    participant AEmbed as aembedding()
    participant GetClient as get_azure_openai_client()
    participant IsV1 as _is_azure_v1_api_version()

    Caller->>Embed: "aembedding=True, api_version=v1"
    Note over Embed: BEFORE FIX: api_version was dropped
    Embed->>AEmbed: "...args, api_version=v1 (fix)"
    AEmbed->>GetClient: "api_version=v1, _is_async=True"
    GetClient->>IsV1: "api_version=v1"
    IsV1-->>GetClient: "True"
    GetClient-->>AEmbed: "AsyncOpenAI (not AsyncAzureOpenAI)"
    AEmbed-->>Caller: "EmbeddingResponse"

Loading

_{Reviews (1): Last reviewed commit: "fix(azure): forward api_version to aembe..." | Re-trigger Greptile}

[greptile-apps]

sys.path depth is one level too deep

The path insert navigates 5 levels up from the file's directory (tests/litellm/llms/azure/), landing in the parent of the repo root rather than the repo root itself. Comparing with tests/litellm/llms/openai_like/test_assemblyai_provider.py — which sits at the same nesting depth and correctly uses "../../../.." (4 levels) — this should also use 4 .. segments:

Suggested change

	sys.path.insert(
	0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../../../../.."))
	)
	sys.path.insert(
	0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../../../.."))
	)

In practice pytest's rootdir detection keeps the tests passing regardless, but the intent is wrong and could confuse future readers or fail in unusual invocation contexts.