fix(azure): preserve content_policy_violation error details from Azure OpenAI

[skylarkoo7]

Relevant issues

Closes #20811

What this PR does

Azure OpenAI returns rich error payloads for content policy violations including inner_error with ResponsibleAIPolicyViolation, content_filter_results, and revised_prompt. These details were being lost and replaced by a generic HTTP 400 response.

Root causes fixed

1. Incomplete azure_error_code extraction: The structured check in exception_mapping_utils.py only examined the top-level body["error"]["code"]. When Azure returns a generic top-level code but nests ResponsibleAIPolicyViolation in inner_error.code, the content-policy path was skipped entirely, falling through to generic BadRequestError handlers that discard error details.

2. Image generation polling path loses structured body: The DALL-E polling path in azure.py stringified the entire error JSON into the message field of AzureOpenAIError without setting the body attribute. This made it impossible for exception_type() to extract structured error details.

3. Overly broad string-based detection: is_azure_content_policy_violation_error() used "invalid_request_error" as a content-policy indicator, which could misclassify regular bad-request errors. Replaced with specific Azure safety-system message patterns.

Changes

File	Change
litellm/litellm_core_utils/exception_mapping_utils.py	Check inner_error.code for ResponsibleAIPolicyViolation when top-level code differs; replace broad string match with specific patterns
litellm/llms/azure/azure.py	Set structured body on AzureOpenAIError in both async and sync DALL-E polling paths
tests/test_litellm/llms/azure/test_azure_exception_mapping.py	Add 5 regression tests covering the exact error payloads from issue #20811

Pre-Submission checklist

• I have Added testing in the tests/litellm/ directory
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Feb 11, 2026 6:44am

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

Greptile Overview

Greptile Summary

This PR improves Azure OpenAI content-policy violation handling, especially for Images/DALL-E flows.

• In litellm_core_utils/exception_mapping_utils.py, Azure content-policy detection now prefers structured error payloads and can treat error.inner_error.code == ResponsibleAIPolicyViolation as a content-policy violation even when the top-level code is generic. It also narrows string-based detection to more specific Azure safety-system phrases.
• In litellm/llms/azure/azure.py, the DALL-E operation polling failure path now raises AzureOpenAIError with a structured body (instead of only a JSON-stringified message), enabling exception_type() to preserve and route rich error details.
• Adds regression tests to ensure these Azure Images payload shapes map to ContentPolicyViolationError and that nested policy fields are preserved.

Confidence Score: 4/5

• Generally safe to merge after addressing small correctness/test hygiene issues.
• The main behavioral changes are narrowly scoped and backed by targeted regression tests, but there are still unhandled Azure error-body shape variants that will deterministically bypass the new inner_error-based detection, plus a stray debug print in tests.
• litellm/litellm_core_utils/exception_mapping_utils.py; tests/test_litellm/llms/azure/test_azure_exception_mapping.py

Important Files Changed

Filename	Overview
litellm/litellm_core_utils/exception_mapping_utils.py	Refines Azure content-policy detection by removing broad substring matching and additionally checking nested inner_error.code for ResponsibleAIPolicyViolation. Main concern: nested-body shape handling is still limited to body['error'].* variants.
litellm/llms/azure/azure.py	Updates Azure DALL-E polling failure handling to preserve structured error payloads by setting body on AzureOpenAIError and using a clearer message extracted from the payload.
tests/test_litellm/llms/azure/test_azure_exception_mapping.py	Adds regression tests for Azure Images content-policy violations and polling error body preservation. Issue: contains a stray debug print() that should be removed to keep tests quiet.

Sequence Diagram

sequenceDiagram
  participant Caller as LiteLLM caller
  participant AzurePy as litellm/llms/azure/azure.py
  participant HTTP as HTTPHandler/httpx
  participant ExMap as exception_mapping_utils.exception_type
  participant AzureMap as AzureOpenAIExceptionMapping

  Caller->>AzurePy: aimage_generation()/sync image generation
  AzurePy->>HTTP: POST submit image generation
  HTTP-->>AzurePy: 202 + operation-location
  loop Poll until done
    AzurePy->>HTTP: GET operation-location
    HTTP-->>AzurePy: {status: succeeded|failed, ...}
  end
  alt status == failed
    AzurePy->>AzurePy: raise AzureOpenAIError(message, body=error_data)
    Caller->>ExMap: exception_type(original_exception)
    ExMap->>ExMap: extract body['error'].code
    ExMap->>ExMap: check body['error'].inner_error.code
    alt content policy detected
      ExMap->>AzureMap: create_content_policy_violation_error(...)
      AzureMap-->>Caller: ContentPolicyViolationError (preserves provider fields)
    else other bad request
      ExMap-->>Caller: BadRequestError
    end
  else status == succeeded
    AzurePy-->>Caller: Image result
  end

Loading

[greptile-apps]

_{3 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Inner error check too narrow

exception_type() only inspects body["error"]["inner_error"] / innererror for ResponsibleAIPolicyViolation. Azure error payloads can also include the inner object under the top-level body (e.g. body["inner_error"]) or use different casing, so those cases will still fall through to BadRequestError and drop structured details. This matters when the exception body mirrors the Images API error without the error wrapper.

Also appears in: tests/test_litellm/llms/azure/test_azure_exception_mapping.py fixtures only cover the body["error"]["inner_error"] and body["innererror"] shapes, so the missing shapes are untested.

[greptile-apps]

Additional Comments (1)

tests/test_litellm/llms/azure/test_azure_exception_mapping.py
Debug print in test

This print("got provider_specific_fields=", ...) will pollute test output (and may fail in setups that treat unexpected stdout as noise). It looks like leftover debugging and should be removed.