# src/agents/system_instructions_rag.py from typing import List, Dict, Optional from src.agents.rag_agent import RAGResponse from src.utils.logger import logger from src.agents.rag_agent import RAGAgent from src.utils.greeting_handler import GreetingHandler class SystemInstructionsRAGAgent(RAGAgent): """RAG Agent with enhanced system instructions for specific use cases""" async def generate_response( self, query: str, conversation_id: Optional[str] = None, temperature: float = 0.7, max_tokens: Optional[int] = None, context_docs: Optional[List[str]] = None ) -> RAGResponse: """Generate response with specific handling for introduction and no-context cases""" try: # First, check if this is an introduction/welcome message query is_introduction = ( "wants support" in query and "This is Introduction" in query and ("A new user with name:" in query or "An old user with name:" in query) ) if is_introduction: # Handle introduction message - no context needed welcome_message = self._handle_contact_query(query) return RAGResponse( response=welcome_message, context_docs=[], sources=[], scores=None ) # Check if this is a greeting (using the singleton GreetingHandler) greeting_handler = GreetingHandler() is_greeting, greeting_response = greeting_handler.is_greeting( query) if is_greeting: logger.info(f"Detected greeting in query: {query}") return RAGResponse( response=greeting_response, context_docs=[], sources=[], scores=None ) # Get conversation history if conversation_id exists conversation_history = [] if conversation_id: try: conversation_history = await self.mongodb.get_recent_messages( conversation_id, limit=self.conversation_manager.max_messages ) # Get relevant history within token limits conversation_history = self.conversation_manager.get_relevant_history( messages=conversation_history, current_query=query ) except Exception as e: logger.warning( f"Error fetching conversation history: {str(e)}") # For all other queries, proceed with context-based response if not context_docs: context_docs, sources, scores = await self.retrieve_context( query, conversation_history=conversation_history ) # Check if we have relevant context has_relevant_context = self._check_context_relevance( query, context_docs or [] ) # If no relevant context found, return the standard message if not has_relevant_context: return RAGResponse( response="Information about this is not available, do you want to inquire about something else?", context_docs=[], sources=[], scores=None ) # Generate response using context and conversation history prompt = self._create_response_prompt( query=query, context_docs=context_docs, conversation_history=conversation_history ) response_text = self.llm.generate( prompt, temperature=temperature, max_tokens=max_tokens ) # Check if the generated response indicates no information cleaned_response = self._clean_response(response_text) if self._is_no_info_response(cleaned_response): return RAGResponse( response="Information about this is not available, do you want to inquire about something else?", context_docs=[], sources=[], scores=None ) return RAGResponse( response=cleaned_response, context_docs=context_docs, sources=sources, scores=scores ) except Exception as e: logger.error(f"Error in SystemInstructionsRAGAgent: {str(e)}") raise def _create_response_prompt( self, query: str, context_docs: List[str], conversation_history: Optional[List[Dict]] = None ) -> str: """Create prompt for generating response from context and conversation history""" # Format context documents formatted_context = '\n\n'.join( f"Context {i+1}:\n{doc.strip()}" for i, doc in enumerate(context_docs) if doc and doc.strip() ) # Format conversation history if available history_context = "" if conversation_history: history_messages = [] # Use last 3 messages for context for msg in conversation_history[-3:]: role = msg.get('role', 'unknown') content = msg.get('content', '') history_messages.append(f"{role.capitalize()}: {content}") if history_messages: history_context = "\nPrevious Conversation:\n" + \ "\n".join(history_messages) return f""" Use the following context and conversation history to provide information about: {query} Context Information: {formatted_context} {history_context} Instructions: 1. Use information from both the context and conversation history 2. If the information is found, provide a direct and concise response 3. Do not make assumptions or add information not present in the context 4. Ensure the response is clear and complete based on available information 5. If you cannot find relevant information about the specific query, respond exactly with: "Information about this is not available, do you want to inquire about something else?" Query: {query} Response:""" def _is_no_info_response(self, response: str) -> bool: """Check if the response indicates no information available""" no_info_indicators = [ "i do not have", "i don't have", "no information", "not available", "could not find", "couldn't find", "cannot find" ] response_lower = response.lower() return any(indicator in response_lower for indicator in no_info_indicators) def _check_context_relevance(self, query: str, context_docs: List[str]) -> bool: """Check if context contains information relevant to the query""" if not context_docs: return False # Extract key terms from query (keeping important words) query_words = query.lower().split() stop_words = {'me', 'a', 'about', 'what', 'is', 'are', 'the', 'in', 'how', 'why', 'when', 'where'} # Remove only basic stop words, keep important terms like "report", "share", etc. query_terms = {word for word in query_words if word not in stop_words} # Add additional relevant terms that might appear in the content related_terms = { 'comprehensive', 'report', 'overview', 'summary', 'details', 'information' } query_terms.update( word for word in query_words if word in related_terms) # Check each context document for relevance for doc in context_docs: if not doc: continue doc_lower = doc.lower() # Consider document relevant if it contains any query terms # or if it starts with common report headers if any(term in doc_lower for term in query_terms) or \ any(header in doc_lower for header in ['overview', 'comprehensive report', 'summary']): return True return False def _handle_contact_query(self, query: str) -> str: """Handle queries from /user/contact endpoint""" try: name_start = query.find('name: "') + 7 name_end = query.find('"', name_start) name = query[name_start:name_end] if name_start > 6 and name_end != -1 else "there" is_returning = ( "An old user with name:" in query and "wants support again" in query ) if is_returning: return f"Welcome back {name}, How can I help you?" return f"Welcome {name}, How can I help you?" except Exception as e: logger.error(f"Error handling contact query: {str(e)}") return "Welcome, How can I help you?" def _clean_response(self, response: str) -> str: """Clean response by removing unwanted phrases""" if not response: return response phrases_to_remove = [ "Based on the context provided,", "According to the documents,", "From the information available,", "I can tell you that", "Let me help you with that", "I understand you're asking about", "To answer your question,", "The documents indicate that", "Based on the available information,", "As per the provided context,", "I would be happy to help you with that", "Let me provide you with information about", "Here's what I found:", "Here's the information you requested:", "According to the provided information,", "Based on the documents,", "The information suggests that", "From what I can see,", "Let me explain", "To clarify,", "It appears that", "I can see that", "Sure,", "Well,", "Based on the given context,", "The available information shows that", "From the context provided,", "The documentation mentions that", "According to the context,", "As shown in the context,", "I apologize," ] cleaned_response = response for phrase in phrases_to_remove: cleaned_response = cleaned_response.replace(phrase, "").strip() cleaned_response = " ".join(cleaned_response.split()) if not cleaned_response: return response if cleaned_response[0].islower(): cleaned_response = cleaned_response[0].upper( ) + cleaned_response[1:] return cleaned_response