Add message moderation in handlers

bot/models/base_handlers.py

@@ -39,6 +39,10 @@ async def get_localized_text(
     async def initialize_context(self, person: Person, context: Context) -> None:
         raise NotImplementedError
 
+    @abstractmethod
+    async def moderate_message(self, message: Message) -> bool:
+        raise NotImplementedError
+
     @property
     def permissions(self) -> List[BasePermission]:
         return self._initialized_permissions
@@ -90,6 +94,10 @@ async def handle(
             return self.get_localized_response(
                 CommandResponse(text="not_authenticated"), context
             )
+        if not await self.moderate_message(message):
+            return self.get_localized_response(
+                CommandResponse(text="message_moderation_failed"), context
+            )
         await self._log_handler_request(person, context, message, args)
         response = await self.get_response(person, context, message, args)
         if response is None:
@@ -106,6 +114,12 @@ async def stream_handle(
                 context,
             )
             return
+        if not await self.moderate_message(message):
+            yield self.get_localized_response(
+                CommandResponse(text="message_moderation_failed"),
+                context,
+            )
+            return
         await self._log_handler_request(person, context, message, args)
         async for chunk in self.stream_get_response(person, context, message, args):
             if chunk is None:

bot/rp_bot/rp_bot_handlers.py

@@ -3,6 +3,7 @@
 
 from .db import DB
 from ai_agent.ai import AI
+from ai_agent.moderation import moderate_user_message
 from .prompt_manager import PromptManager
 from .auth import Auth
 from .localizer import Localizer
@@ -13,7 +14,7 @@
     BaseMessageHandler,
 )
 from ..models.base_auth import BasePermission
-from ..models.handlers_input import Person, Context
+from ..models.handlers_input import Person, Context, Message
 
 
 class RPBotHandlerMixin(ABC):
@@ -47,6 +48,9 @@ def get_initialized_permissions(self) -> List[BasePermission]:
             result.append(permission_class(self.auth))
         return result
 
+    async def moderate_message(self, message: Message) -> bool:
+        return moderate_user_message(message)
+
     async def get_localized_text(
         self, text: str, kwargs: Optional[dict] = None, context: Optional[Context] = None
     ) -> Optional[str]:

config/localizer_translations.yaml

@@ -83,6 +83,8 @@ LocalizerTranslations:
     english: "{response_text}"
   streaming_message_response:
     english: "{response_text}"
+  message_moderation_failed:
+    english: "Message moderation failed"
   usage_limit_exceeded:
     english: "Usage limit exceeded for {user_handle} with {user_limit} limit"
   content_not_acceptable: