BerriAI/litellm · error · ValueError

api_key is None. Please set AZURE_AI_API_KEY or dynamically

Error message

api_key is None. Please set AZURE_AI_API_KEY or dynamically via `api_key` param, to make the request.

What it means

Twin guard of the api_base check in the Azure AI image-embeddings handler: after confirming api_base, it requires an api_key for the Authorization: Bearer header and raises ValueError if none, pointing at AZURE_AI_API_KEY or the api_key parameter. Raised client-side before the request is sent.

Source

Thrown at litellm/llms/azure_ai/embed/handler.py:106

    def image_embedding(
        self,
        model: str,
        data: ImageEmbeddingRequest,
        timeout: float,
        logging_obj,
        model_response: EmbeddingResponse,
        optional_params: dict,
        api_key: str | None,
        api_base: str | None,
        client: HTTPHandler | AsyncHTTPHandler | None = None,
    ):
        if api_base is None:
            raise ValueError(
                "api_base is None. Please set AZURE_AI_API_BASE or dynamically via `api_base` param, to make the request."
            )
        if api_key is None:
            raise ValueError(
                "api_key is None. Please set AZURE_AI_API_KEY or dynamically via `api_key` param, to make the request."
            )

        if client is None or not isinstance(client, HTTPHandler):
            client = HTTPHandler(timeout=timeout, concurrent_limit=1)

        url: Final = f"{api_base}/images/embeddings"

        response: Final = client.post(
            url=url,
            json=data,
            headers={"Authorization": f"Bearer {api_key}"},
        )

        embedding_response: Final = response.json()
        embedding_headers: Final = dict(response.headers)
        returned_response: Final[EmbeddingResponse] = convert_to_model_response_object(
            response_object=embedding_response,

View on GitHub (pinned to 6c2dcb801b)

Solutions

Pass api_key=<foundry key> on the call.
Or export AZURE_AI_API_KEY in the litellm process environment.
In proxy config use api_key: os.environ/AZURE_AI_API_KEY so it resolves at startup and fails fast.
Copy the key from the Foundry project's 'Keys' page — it is the project key, not the Azure subscription key.

Example fix

# before
litellm.embedding(model='azure_ai/img-embed', input=[b64], api_base=base)

# after
litellm.embedding(model='azure_ai/img-embed', input=[b64], api_base=base,
                  api_key=os.environ['AZURE_AI_API_KEY'])

Defensive patterns

Strategy: validation

Validate before calling

import os

def foundry_key() -> str:
    key = os.getenv('AZURE_AI_API_KEY')
    if not key:
        raise RuntimeError('AZURE_AI_API_KEY is required for azure_ai embeddings')
    return key

Try / catch

try:
    litellm.embedding(model='azure_ai/img', input=imgs, api_key=foundry_key())
except ValueError as e:
    if 'AZURE_AI_API_KEY' in str(e):
        raise ConfigurationError('missing Foundry API key') from e
    raise

Prevention

Resolve api_key: os.environ/AZURE_AI_API_KEY in proxy config so absence fails at startup.
Distinguish AZURE_API_KEY from AZURE_AI_API_KEY in your secret naming.
Add a pre-flight authenticated call (list models) in health checks to catch bad keys early.

When it happens

Trigger: Azure AI embedding call with api_base set but no api_key and no AZURE_AI_API_KEY env var; key passed as azure_ad_token or under a different name; env var unset in containerized deployments.

Common situations: Secrets moved to a vault and the AZURE_AI_API_KEY injection forgotten in CI; using a Key Vault reference name but the proxy never resolves it; local .env not loaded in the deployed image.

Related errors

AI-assisted analysis of BerriAI/litellm@6c2dcb801b (2026-08-15). Data as JSON: /api/errors/33faf6f4aad92f53. Report an issue: GitHub.