BerriAI/litellm · error · AzureOpenAIError

{str(e)}

Error message

{str(e)}

What it means

This is the catch-all handler at the end of Azure text completion: any exception that is not itself AzureOpenAIError is converted into AzureOpenAIError, preserving the original status_code/headers attributes if present (defaulting to 500). The message is str(e) of the underlying error — typically an openai.APIStatusError from the Azure OpenAI service.

Source

Thrown at litellm/llms/azure/completion/handler.py:185

                    additional_args={
                        "headers": headers,
                        "api_version": api_version,
                        "api_base": api_base,
                    },
                )
                return openai_text_completion_config.convert_to_chat_model_response_object(
                    response_object=TextCompletionResponse(**stringified_response),
                    model_response_object=model_response,
                )
        except AzureOpenAIError as e:
            raise e
        except Exception as e:
            status_code: Final = getattr(e, "status_code", 500)
            error_headers = getattr(e, "headers", None)
            error_response: Final = getattr(e, "response", None)
            if error_headers is None and error_response:
                error_headers = getattr(error_response, "headers", None)
            raise AzureOpenAIError(status_code=status_code, message=str(e), headers=error_headers)

    async def acompletion(
        self,
        api_key: str | None,
        api_version: str,
        model: str,
        api_base: str,
        data: dict,
        timeout: Any,
        model_response: ModelResponse,
        logging_obj: Any,
        max_retries: int,
        azure_ad_token: str | None = None,
        client=None,  # this is the AsyncAzureOpenAI
        litellm_params: dict = {},
    ):
        response = None
        try:

View on GitHub (pinned to 6c2dcb801b)

Solutions

1. Inspect the wrapped error's status_code and message — they mirror the upstream Azure/OpenAI error and its headers (retry-after for 429s).
2. 401/403: verify AZURE_API_KEY and the deployment's access settings; 404: verify the deployment name after 'azure/'.
3. 429: honor retry-after, reduce RPM/TPM, or enable LiteLLM router retries/cooldowns.
4. Timeouts: pass a larger timeout= to the call and check network egress to <resource>.openai.azure.com.

Example fix

# before
resp = litellm.text_completion(model="azure/dep", prompt=txt, timeout=5)  # ReadTimeout -> 500 wrap

# after
resp = litellm.text_completion(model="azure/dep", prompt=txt, timeout=120, max_retries=3)

from litellm.exceptions import AzureOpenAIError

try:
    resp = litellm.text_completion(model="azure/dep", prompt=txt)
except AzureOpenAIError as e:
    if e.status_code == 429:
        wait = int(e.headers.get("retry-after", "5")) if e.headers else 5
        time.sleep(wait)
        resp = litellm.text_completion(model="azure/dep", prompt=txt)
    elif e.status_code in (401, 403):
        raise AuthError(str(e)) from e
    elif e.status_code == 404:
        raise ConfigError("Deployment not found — check model name") from e
    else:
        raise

When it happens

Trigger: The Azure OpenAI service returned 4xx/5xx (401 bad key, 404 unknown deployment, 429 quota, 500s) and the OpenAI SDK raised; network timeouts (httpx.ConnectTimeout/ReadTimeout); JSON decode errors on truncated responses.

Common situations: Wrong api_key or expired key (401); deployment name in the model string does not exist in the resource (404); hitting token-per-minute rate limits (429); regional outages; short client timeouts on long completions.

Related errors

• raw_response.text
• Azure Document Intelligence analysis failed: {error_msg}
• Azure Document Intelligence analysis failed with status: {op
• error_msg (upstream response error message)
• raw_response.text

AI-assisted analysis of BerriAI/litellm@6c2dcb801b (2026-08-15). Data as JSON: /api/errors/5b1881a054d8ac13. Report an issue: GitHub.