feat: add hosted moderation (#1158)

api/core/model_providers/model_factory.py

@@ -8,6 +8,7 @@ from core.model_providers.models.base import BaseProviderModel
 from core.model_providers.models.embedding.base import BaseEmbedding
 from core.model_providers.models.entity.model_params import ModelKwargs, ModelType
 from core.model_providers.models.llm.base import BaseLLM
+from core.model_providers.models.moderation.base import BaseModeration
 from core.model_providers.models.speech2text.base import BaseSpeech2Text
 from extensions.ext_database import db
 from models.provider import TenantDefaultModel
@@ -180,7 +181,7 @@ class ModelFactory:
     def get_moderation_model(cls,
                              tenant_id: str,
                              model_provider_name: str,
-                             model_name: str) -> Optional[BaseProviderModel]:
+                             model_name: str) -> Optional[BaseModeration]:
         """
         get moderation model.
 

api/core/model_providers/models/llm/base.py

@@ -10,6 +10,7 @@ from langchain.memory.chat_memory import BaseChatMemory
 from langchain.schema import LLMResult, SystemMessage, AIMessage, HumanMessage, BaseMessage, ChatGeneration
 
 from core.callback_handler.std_out_callback_handler import DifyStreamingStdOutCallbackHandler, DifyStdOutCallbackHandler
+from core.helper import moderation
 from core.model_providers.models.base import BaseProviderModel
 from core.model_providers.models.entity.message import PromptMessage, MessageType, LLMRunResult, to_prompt_messages
 from core.model_providers.models.entity.model_params import ModelType, ModelKwargs, ModelMode, ModelKwargsRules
@@ -116,6 +117,15 @@ class BaseLLM(BaseProviderModel):
         :param callbacks:
         :return:
         """
+        moderation_result = moderation.check_moderation(
+            self.model_provider,
+            "\n".join([message.content for message in messages])
+        )
+
+        if not moderation_result:
+            kwargs['fake_response'] = "I apologize for any confusion, " \
+                                      "but I'm an AI assistant to be helpful, harmless, and honest."
+
         if self.deduct_quota:
             self.model_provider.check_quota_over_limit()
 

api/core/model_providers/models/moderation/base.py

@@ -0,0 +1,29 @@
+from abc import abstractmethod
+from typing import Any
+
+from core.model_providers.models.base import BaseProviderModel
+from core.model_providers.models.entity.model_params import ModelType
+from core.model_providers.providers.base import BaseModelProvider
+
+
+class BaseModeration(BaseProviderModel):
+    name: str
+    type: ModelType = ModelType.MODERATION
+
+    def __init__(self, model_provider: BaseModelProvider, client: Any, name: str):
+        super().__init__(model_provider, client)
+        self.name = name
+
+    def run(self, text: str) -> bool:
+        try:
+            return self._run(text)
+        except Exception as ex:
+            raise self.handle_exceptions(ex)
+
+    @abstractmethod
+    def _run(self, text: str) -> bool:
+        raise NotImplementedError
+
+    @abstractmethod
+    def handle_exceptions(self, ex: Exception) -> Exception:
+        raise NotImplementedError

api/core/model_providers/models/moderation/openai_moderation.py

@@ -4,29 +4,35 @@ import openai
 
 from core.model_providers.error import LLMBadRequestError, LLMAPIConnectionError, LLMAPIUnavailableError, \
     LLMRateLimitError, LLMAuthorizationError
-from core.model_providers.models.base import BaseProviderModel
-from core.model_providers.models.entity.model_params import ModelType
+from core.model_providers.models.moderation.base import BaseModeration
 from core.model_providers.providers.base import BaseModelProvider
 
-DEFAULT_AUDIO_MODEL = 'whisper-1'
+DEFAULT_MODEL = 'whisper-1'
 
 
-class OpenAIModeration(BaseProviderModel):
-    type: ModelType = ModelType.MODERATION
+class OpenAIModeration(BaseModeration):
 
     def __init__(self, model_provider: BaseModelProvider, name: str):
-        super().__init__(model_provider, openai.Moderation)
+        super().__init__(model_provider, openai.Moderation, name)
 
-    def run(self, text):
+    def _run(self, text: str) -> bool:
         credentials = self.model_provider.get_model_credentials(
-            model_name=DEFAULT_AUDIO_MODEL,
+            model_name=self.name,
             model_type=self.type
         )
 
-        try:
-            return self._client.create(input=text, api_key=credentials['openai_api_key'])
-        except Exception as ex:
-            raise self.handle_exceptions(ex)
+        # 2000 text per chunk
+        length = 2000
+        chunks = [text[i:i + length] for i in range(0, len(text), length)]
+
+        moderation_result = self._client.create(input=chunks,
+                                                api_key=credentials['openai_api_key'])
+
+        for result in moderation_result.results:
+            if result['flagged'] is True:
+                return False
+
+        return True
 
     def handle_exceptions(self, ex: Exception) -> Exception:
         if isinstance(ex, openai.error.InvalidRequestError):