feat(router.py): support content policy fallbacks

Closes https://github.com/BerriAI/litellm/issues/2632

litellm/__init__.py

@@ -240,6 +240,7 @@ num_retries: Optional[int] = None  # per model endpoint
 default_fallbacks: Optional[List] = None
 fallbacks: Optional[List] = None
 context_window_fallbacks: Optional[List] = None
+content_policy_fallbacks: Optional[List] = None
 allowed_fails: int = 0
 num_retries_per_request: Optional[int] = (
     None  # for the request overall (incl. fallbacks + model retries)

litellm/exceptions.py

@@ -324,7 +324,7 @@ class ContextWindowExceededError(BadRequestError):  # type: ignore
         message,
         model,
         llm_provider,
-        response: httpx.Response,
+        response: Optional[httpx.Response] = None,
         litellm_debug_info: Optional[str] = None,
     ):
         self.status_code = 400
@@ -332,11 +332,13 @@ class ContextWindowExceededError(BadRequestError):  # type: ignore
         self.model = model
         self.llm_provider = llm_provider
         self.litellm_debug_info = litellm_debug_info
+        request = httpx.Request(method="POST", url="https://api.openai.com/v1")
+        self.response = response or httpx.Response(status_code=400, request=request)
         super().__init__(
             message=self.message,
             model=self.model,  # type: ignore
             llm_provider=self.llm_provider,  # type: ignore
-            response=response,
+            response=self.response,
             litellm_debug_info=self.litellm_debug_info,
         )  # Call the base class constructor with the parameters it needs
 
@@ -407,7 +409,7 @@ class ContentPolicyViolationError(BadRequestError):  # type: ignore
         message,
         model,
         llm_provider,
-        response: httpx.Response,
+        response: Optional[httpx.Response] = None,
         litellm_debug_info: Optional[str] = None,
     ):
         self.status_code = 400
@@ -415,11 +417,13 @@ class ContentPolicyViolationError(BadRequestError):  # type: ignore
         self.model = model
         self.llm_provider = llm_provider
         self.litellm_debug_info = litellm_debug_info
+        request = httpx.Request(method="POST", url="https://api.openai.com/v1")
+        self.response = response or httpx.Response(status_code=500, request=request)
         super().__init__(
             message=self.message,
             model=self.model,  # type: ignore
             llm_provider=self.llm_provider,  # type: ignore
-            response=response,
+            response=self.response,
             litellm_debug_info=self.litellm_debug_info,
         )  # Call the base class constructor with the parameters it needs
 

litellm/main.py

@@ -401,6 +401,7 @@ def mock_completion(
     stream: Optional[bool] = False,
     mock_response: Union[str, Exception] = "This is a mock request",
     logging=None,
+    custom_llm_provider=None,
     **kwargs,
 ):
     """
@@ -438,7 +439,7 @@ def mock_completion(
             raise litellm.APIError(
                 status_code=getattr(mock_response, "status_code", 500),  # type: ignore
                 message=getattr(mock_response, "text", str(mock_response)),
-                llm_provider=getattr(mock_response, "llm_provider", "openai"),  # type: ignore
+                llm_provider=getattr(mock_response, "llm_provider", custom_llm_provider or "openai"),  # type: ignore
                 model=model,  # type: ignore
                 request=httpx.Request(method="POST", url="https://api.openai.com/v1/"),
             )
@@ -907,6 +908,7 @@ def completion(
                 logging=logging,
                 acompletion=acompletion,
                 mock_delay=kwargs.get("mock_delay", None),
+                custom_llm_provider=custom_llm_provider,
             )
         if custom_llm_provider == "azure":
             # azure configs

litellm/router.py

@@ -108,6 +108,7 @@ class Router:
         ] = None,  # generic fallbacks, works across all deployments
         fallbacks: List = [],
         context_window_fallbacks: List = [],
+        content_policy_fallbacks: List = [],
         model_group_alias: Optional[dict] = {},
         enable_pre_call_checks: bool = False,
         retry_after: int = 0,  # min time to wait before retrying a failed request
@@ -311,6 +312,12 @@ class Router:
         self.context_window_fallbacks = (
             context_window_fallbacks or litellm.context_window_fallbacks
         )
+
+        _content_policy_fallbacks = (
+            content_policy_fallbacks or litellm.content_policy_fallbacks
+        )
+        self.validate_fallbacks(fallback_param=_content_policy_fallbacks)
+        self.content_policy_fallbacks = _content_policy_fallbacks
         self.total_calls: defaultdict = defaultdict(
             int
         )  # dict to store total calls made to each model
@@ -1998,6 +2005,9 @@ class Router:
         context_window_fallbacks = kwargs.get(
             "context_window_fallbacks", self.context_window_fallbacks
         )
+        content_policy_fallbacks = kwargs.get(
+            "content_policy_fallbacks", self.content_policy_fallbacks
+        )
         try:
             if mock_testing_fallbacks is not None and mock_testing_fallbacks == True:
                 raise Exception(
@@ -2016,7 +2026,10 @@ class Router:
                 if (
                     hasattr(e, "status_code")
                     and e.status_code == 400  # type: ignore
-                    and not isinstance(e, litellm.ContextWindowExceededError)
+                    and not (
+                        isinstance(e, litellm.ContextWindowExceededError)
+                        or isinstance(e, litellm.ContentPolicyViolationError)
+                    )
                 ):  # don't retry a malformed request
                     raise e
                 if (
@@ -2034,6 +2047,39 @@ class Router:
                     if fallback_model_group is None:
                         raise original_exception
 
+                    for mg in fallback_model_group:
+                        """
+                        Iterate through the model groups and try calling that deployment
+                        """
+                        try:
+                            kwargs["model"] = mg
+                            kwargs.setdefault("metadata", {}).update(
+                                {"model_group": mg}
+                            )  # update model_group used, if fallbacks are done
+                            response = await self.async_function_with_retries(
+                                *args, **kwargs
+                            )
+                            verbose_router_logger.info(
+                                "Successful fallback b/w models."
+                            )
+                            return response
+                        except Exception as e:
+                            pass
+                elif (
+                    isinstance(e, litellm.ContentPolicyViolationError)
+                    and content_policy_fallbacks is not None
+                ):
+                    fallback_model_group = None
+                    for (
+                        item
+                    ) in content_policy_fallbacks:  # [{"gpt-3.5-turbo": ["gpt-4"]}]
+                        if list(item.keys())[0] == model_group:
+                            fallback_model_group = item[model_group]
+                            break
+
+                    if fallback_model_group is None:
+                        raise original_exception
+
                     for mg in fallback_model_group:
                         """
                         Iterate through the model groups and try calling that deployment
@@ -2114,6 +2160,9 @@ class Router:
         context_window_fallbacks = kwargs.pop(
             "context_window_fallbacks", self.context_window_fallbacks
         )
+        content_policy_fallbacks = kwargs.pop(
+            "content_policy_fallbacks", self.content_policy_fallbacks
+        )
 
         num_retries = kwargs.pop("num_retries")
 
@@ -2141,6 +2190,7 @@ class Router:
                 healthy_deployments=_healthy_deployments,
                 context_window_fallbacks=context_window_fallbacks,
                 regular_fallbacks=fallbacks,
+                content_policy_fallbacks=content_policy_fallbacks,
             )
 
             # decides how long to sleep before retry
@@ -2206,10 +2256,12 @@ class Router:
         error: Exception,
         healthy_deployments: Optional[List] = None,
         context_window_fallbacks: Optional[List] = None,
+        content_policy_fallbacks: Optional[List] = None,
         regular_fallbacks: Optional[List] = None,
     ):
         """
         1. raise an exception for ContextWindowExceededError if context_window_fallbacks is not None
+        2. raise an exception for ContentPolicyViolationError if content_policy_fallbacks is not None
 
         2. raise an exception for RateLimitError if
             - there are no fallbacks
@@ -2219,13 +2271,19 @@ class Router:
         if healthy_deployments is not None and isinstance(healthy_deployments, list):
             _num_healthy_deployments = len(healthy_deployments)
 
-        ### CHECK IF RATE LIMIT / CONTEXT WINDOW ERROR w/ fallbacks available / Bad Request Error
+        ### CHECK IF RATE LIMIT / CONTEXT WINDOW ERROR / CONTENT POLICY VIOLATION ERROR w/ fallbacks available / Bad Request Error
         if (
             isinstance(error, litellm.ContextWindowExceededError)
             and context_window_fallbacks is not None
         ):
             raise error
 
+        if (
+            isinstance(error, litellm.ContentPolicyViolationError)
+            and content_policy_fallbacks is not None
+        ):
+            raise error
+
         # Error we should only retry if there are other deployments
         if isinstance(error, openai.RateLimitError):
             if (
@@ -2256,6 +2314,9 @@ class Router:
         context_window_fallbacks = kwargs.get(
             "context_window_fallbacks", self.context_window_fallbacks
         )
+        content_policy_fallbacks = kwargs.get(
+            "content_policy_fallbacks", self.content_policy_fallbacks
+        )
         try:
             if mock_testing_fallbacks is not None and mock_testing_fallbacks == True:
                 raise Exception(
@@ -2271,7 +2332,10 @@ class Router:
                 if (
                     hasattr(e, "status_code")
                     and e.status_code == 400  # type: ignore
-                    and not isinstance(e, litellm.ContextWindowExceededError)
+                    and not (
+                        isinstance(e, litellm.ContextWindowExceededError)
+                        or isinstance(e, litellm.ContentPolicyViolationError)
+                    )
                 ):  # don't retry a malformed request
                     raise e
 
@@ -2294,6 +2358,37 @@ class Router:
                     if fallback_model_group is None:
                         raise original_exception
 
+                    for mg in fallback_model_group:
+                        """
+                        Iterate through the model groups and try calling that deployment
+                        """
+                        try:
+                            ## LOGGING
+                            kwargs = self.log_retry(kwargs=kwargs, e=original_exception)
+                            kwargs["model"] = mg
+                            kwargs.setdefault("metadata", {}).update(
+                                {"model_group": mg}
+                            )  # update model_group used, if fallbacks are done
+                            response = self.function_with_fallbacks(*args, **kwargs)
+                            return response
+                        except Exception as e:
+                            pass
+                elif (
+                    isinstance(e, litellm.ContentPolicyViolationError)
+                    and content_policy_fallbacks is not None
+                ):
+                    fallback_model_group = None
+
+                    for (
+                        item
+                    ) in content_policy_fallbacks:  # [{"gpt-3.5-turbo": ["gpt-4"]}]
+                        if list(item.keys())[0] == model_group:
+                            fallback_model_group = item[model_group]
+                            break
+
+                    if fallback_model_group is None:
+                        raise original_exception
+
                     for mg in fallback_model_group:
                         """
                         Iterate through the model groups and try calling that deployment
@@ -2400,6 +2495,9 @@ class Router:
         context_window_fallbacks = kwargs.pop(
             "context_window_fallbacks", self.context_window_fallbacks
         )
+        content_policy_fallbacks = kwargs.pop(
+            "content_policy_fallbacks", self.content_policy_fallbacks
+        )
 
         try:
             # if the function call is successful, no exception will be raised and we'll break out of the loop
@@ -2419,6 +2517,7 @@ class Router:
                 healthy_deployments=_healthy_deployments,
                 context_window_fallbacks=context_window_fallbacks,
                 regular_fallbacks=fallbacks,
+                content_policy_fallbacks=content_policy_fallbacks,
             )
 
             # decides how long to sleep before retry

litellm/tests/test_router_fallbacks.py

@@ -1109,3 +1109,59 @@ async def test_client_side_fallbacks_list(sync_mode):
 
     assert isinstance(response, litellm.ModelResponse)
     assert response.model is not None and response.model == "gpt-4o"
+
+
+@pytest.mark.parametrize("sync_mode", [True, False])
+@pytest.mark.asyncio
+async def test_router_content_policy_fallbacks(sync_mode):
+    os.environ["LITELLM_LOG"] = "DEBUG"
+    router = Router(
+        model_list=[
+            {
+                "model_name": "claude-2",
+                "litellm_params": {
+                    "model": "claude-2",
+                    "api_key": "",
+                    "mock_response": Exception("content filtering policy"),
+                },
+            },
+            {
+                "model_name": "my-fallback-model",
+                "litellm_params": {
+                    "model": "claude-2",
+                    "api_key": "",
+                    "mock_response": "This works!",
+                },
+            },
+            {
+                "model_name": "my-general-model",
+                "litellm_params": {
+                    "model": "claude-2",
+                    "api_key": "",
+                    "mock_response": Exception("Should not have called this."),
+                },
+            },
+            {
+                "model_name": "my-context-window-model",
+                "litellm_params": {
+                    "model": "claude-2",
+                    "api_key": "",
+                    "mock_response": Exception("Should not have called this."),
+                },
+            },
+        ],
+        content_policy_fallbacks=[{"claude-2": ["my-fallback-model"]}],
+        fallbacks=[{"claude-2": ["my-general-model"]}],
+        context_window_fallbacks=[{"claude-2": ["my-context-window-model"]}],
+    )
+
+    if sync_mode is True:
+        response = router.completion(
+            model="claude-2",
+            messages=[{"role": "user", "content": "Hey, how's it going?"}],
+        )
+    else:
+        response = await router.acompletion(
+            model="claude-2",
+            messages=[{"role": "user", "content": "Hey, how's it going?"}],
+        )

litellm/utils.py

@@ -3436,6 +3436,7 @@ def client(original_function):
                     isinstance(e, litellm.exceptions.ContextWindowExceededError)
                     and context_window_fallback_dict
                     and model in context_window_fallback_dict
+                    and not _is_litellm_router_call
                 ):
                     if len(args) > 0:
                         args[0] = context_window_fallback_dict[model]
@@ -8637,32 +8638,33 @@ def exception_type(
                         ),
                     )
             elif custom_llm_provider == "anthropic":  # one of the anthropics
-                if hasattr(original_exception, "message"):
+                if "prompt is too long" in error_str or "prompt: length" in error_str:
-                    if (
-                        "prompt is too long" in original_exception.message
-                        or "prompt: length" in original_exception.message
-                    ):
                     exception_mapping_worked = True
                     raise ContextWindowExceededError(
-                            message=original_exception.message,
+                        message=error_str,
                         model=model,
                         llm_provider="anthropic",
-                            response=original_exception.response,
                     )
-                    if "Invalid API Key" in original_exception.message:
+                if "Invalid API Key" in error_str:
                     exception_mapping_worked = True
                     raise AuthenticationError(
-                            message=original_exception.message,
+                        message=error_str,
+                        model=model,
+                        llm_provider="anthropic",
+                    )
+                if "content filtering policy" in error_str:
+                    exception_mapping_worked = True
+                    raise ContentPolicyViolationError(
+                        message=error_str,
                         model=model,
                         llm_provider="anthropic",
-                            response=original_exception.response,
                     )
                 if hasattr(original_exception, "status_code"):
                     print_verbose(f"status_code: {original_exception.status_code}")
                     if original_exception.status_code == 401:
                         exception_mapping_worked = True
                         raise AuthenticationError(
-                            message=f"AnthropicException - {original_exception.message}",
+                            message=f"AnthropicException - {error_str}",
                             llm_provider="anthropic",
                             model=model,
                             response=original_exception.response,
@@ -8673,7 +8675,7 @@ def exception_type(
                     ):
                         exception_mapping_worked = True
                         raise BadRequestError(
-                            message=f"AnthropicException - {original_exception.message}",
+                            message=f"AnthropicException - {error_str}",
                             model=model,
                             llm_provider="anthropic",
                             response=original_exception.response,
@@ -8681,14 +8683,14 @@ def exception_type(
                     elif original_exception.status_code == 408:
                         exception_mapping_worked = True
                         raise Timeout(
-                            message=f"AnthropicException - {original_exception.message}",
+                            message=f"AnthropicException - {error_str}",
                             model=model,
                             llm_provider="anthropic",
                         )
                     elif original_exception.status_code == 429:
                         exception_mapping_worked = True
                         raise RateLimitError(
-                            message=f"AnthropicException - {original_exception.message}",
+                            message=f"AnthropicException - {error_str}",
                             llm_provider="anthropic",
                             model=model,
                             response=original_exception.response,
@@ -8697,7 +8699,7 @@ def exception_type(
                         exception_mapping_worked = True
                         raise APIError(
                             status_code=500,
-                            message=f"AnthropicException - {original_exception.message}. Handle with `litellm.APIError`.",
+                            message=f"AnthropicException - {error_str}. Handle with `litellm.APIError`.",
                             llm_provider="anthropic",
                             model=model,
                             request=original_exception.request,