# DEPENDENCIES import time import asyncio from typing import Dict from typing import List from typing import Optional from datetime import datetime from typing import AsyncGenerator from config.models import PromptType from config.models import LLMProvider from config.models import QueryRequest from config.models import QueryResponse from config.models import ChunkWithScore from config.settings import get_settings from config.logging_config import get_logger from utils.error_handler import handle_errors from generation.llm_client import get_llm_client from utils.error_handler import ResponseGenerationError from generation.prompt_builder import get_prompt_builder from retrieval.hybrid_retriever import get_hybrid_retriever from generation.query_classifier import get_query_classifier from generation.general_responder import get_general_responder from generation.citation_formatter import get_citation_formatter from generation.temperature_controller import get_temperature_controller # Setup Settings and Logging settings = get_settings() logger = get_logger(__name__) class ResponseGenerator: """ Main orchestrator for RAG response generation with LLM-based intelligent query routing Handles both: 1. Generic/conversational queries (greetings, system info, general knowledge) 2. Document-based RAG queries (retrieval + generation) Pipeline: Query → LLM Classifier → Route to (General LLM | RAG Pipeline) → Response """ def __init__(self, provider: LLMProvider = None, model_name: str = None): """ Initialize response generator with LLM-based query routing capabilities Arguments: ---------- provider { LLMProvider } : LLM provider (Ollama/OpenAI) model_name { str } : Model name to use """ self.logger = logger self.settings = get_settings() # Auto-detect provider for HF Spaces if provider is None: if (self.settings.IS_HF_SPACE and not self.settings.OLLAMA_ENABLED): if (self.settings.USE_OPENAI and self.settings.OPENAI_API_KEY): provider = LLMProvider.OPENAI model_name = model_name or self.settings.OPENAI_MODEL logger.info("Auto-detected: Using OpenAI") else: raise ValueError("No LLM provider configured for HF Space. Set OPENAI_API_KEY in Space secrets.") else: # Local development - use Ollama provider = LLMProvider.OLLAMA self.provider = provider self.model_name = model_name # Initialize components self.llm_client = get_llm_client(provider = self.provider, model_name = self.model_name, ) # Query routing components (NOW USES LLM FOR CLASSIFICATION) self.query_classifier = get_query_classifier(provider = self.provider, model_name = self.model_name, ) self.general_responder = get_general_responder(provider = self.provider, model_name = self.model_name, ) # RAG components self.hybrid_retriever = get_hybrid_retriever() self.prompt_builder = get_prompt_builder(model_name = self.model_name) self.citation_formatter = get_citation_formatter() self.temperature_controller = get_temperature_controller() # Statistics self.generation_count = 0 self.total_generation_time = 0.0 self.general_query_count = 0 self.rag_query_count = 0 self.logger.info(f"Initialized ResponseGenerator with LLM-Based Query Routing: provider={self.provider.value}, model={self.model_name}") @handle_errors(error_type = ResponseGenerationError, log_error = True, reraise = True) async def generate_response(self, request: QueryRequest, conversation_history: List[Dict] = None, has_documents: bool = True) -> QueryResponse: """ Generate response with LLM-based intelligent query routing Arguments: ---------- request { QueryRequest } : Query request object conversation_history { list } : Previous conversation messages has_documents { bool } : Whether documents are available in the system Returns: -------- { QueryResponse } : Complete query response """ start_time = time.time() self.logger.info(f"Processing query: '{request.query[:100]}...'") try: # Classify query using LLM classification = await self.query_classifier.classify(query = request.query, has_documents = has_documents, ) self.logger.info(f"Query classified as: {classification['type']} (confidence: {classification['confidence']:.2f}, LLM-based: {classification.get('is_llm_classified', False)})") self.logger.debug(f"Classification reason: {classification['reason']}") # Route based on classification if (classification['suggested_action'] == 'respond_with_general_llm'): # Handle as general query response = await self._handle_general_query(request = request, classification = classification, start_time = start_time, history = conversation_history, ) self.general_query_count += 1 return response elif (classification['suggested_action'] == 'respond_with_rag'): # Handle as RAG query response = await self._handle_rag_query(request = request, classification = classification, start_time = start_time, ) self.rag_query_count += 1 return response else: # Default to RAG if unclear self.logger.info("Unclear classification - defaulting to RAG...") try: response = await self._handle_rag_query(request = request, classification = classification, start_time = start_time, allow_fallback = True, ) # If no results from RAG, fall back to general if ((not response.sources) or (len(response.sources) == 0)): self.logger.info("No RAG results - falling back to general response") response = await self._handle_general_query(request = request, classification = classification, start_time = start_time, history = conversation_history, ) self.general_query_count += 1 else: self.rag_query_count += 1 return response except Exception as e: self.logger.warning(f"RAG attempt failed, falling back to general: {e}") response = await self._handle_general_query(request = request, classification = classification, start_time = start_time, history = conversation_history, ) self.general_query_count += 1 return response except Exception as e: self.logger.error(f"Response generation failed: {repr(e)}", exc_info = True) raise ResponseGenerationError(f"Response generation failed: {repr(e)}") async def _handle_general_query(self, request: QueryRequest, classification: Dict, start_time: float, history: List[Dict] = None) -> QueryResponse: """ Handle general/conversational queries without RAG Arguments: ---------- request { QueryRequest } : Original request classification { dict } : Classification result start_time { float } : Start timestamp history { list } : Conversation history Returns: -------- { QueryResponse } : Response without RAG """ self.logger.debug("Handling as general query...") # Use general responder general_response = await self.general_responder.respond(query = request.query, conversation_history = history, ) answer = general_response.get("answer", "I'm here to help! Please let me know how I can assist you.") total_time = (time.time() - start_time) * 1000 # Create QueryResponse object response = QueryResponse(query = request.query, answer = answer, sources = [], # No sources for general queries retrieval_time_ms = 0.0, generation_time_ms = total_time, total_time_ms = total_time, tokens_used = general_response.get("tokens_used", {"input": 0, "output": 0, "total": 0}), model_used = self.model_name, timestamp = datetime.now(), ) # Add metadata about query type if request.include_metrics: response.metrics = {"query_type" : "general", "classification" : classification['type'], "confidence" : classification['confidence'], "requires_rag" : False, "conversation_mode" : True, "llm_classified" : classification.get('is_llm_classified', False), } self.logger.info(f"General response generated in {total_time:.0f}ms") return response async def _handle_rag_query(self, request: QueryRequest, classification: Dict, start_time: float, allow_fallback: bool = False) -> QueryResponse: """ Handle RAG-based queries with document retrieval Arguments: ---------- request { QueryRequest } : Original request classification { dict } : Classification result start_time { float } : Start timestamp allow_fallback { bool } : Whether to allow fallback to general Returns: -------- { QueryResponse } : RAG response """ self.logger.debug("Handling as RAG query...") try: # Retrieve relevant context self.logger.debug("Retrieving context...") retrieval_start = time.time() retrieval_result = self.hybrid_retriever.retrieve_with_context(query = request.query, top_k = request.top_k or self.settings.TOP_K_RETRIEVE, enable_reranking = request.enable_reranking, include_citations = request.include_sources, ) retrieval_time = (time.time() - retrieval_start) * 1000 chunks = retrieval_result["chunks"] context = retrieval_result["context"] if not chunks: self.logger.warning("No relevant context found for query") if allow_fallback: # Return empty response to trigger fallback return QueryResponse(query = request.query, answer = "", sources = [], retrieval_time_ms = retrieval_time, generation_time_ms = 0.0, total_time_ms = retrieval_time, tokens_used = {"input": 0, "output": 0, "total": 0}, model_used = self.model_name, timestamp = datetime.now(), ) else: return self._create_no_results_response(request = request, retrieval_time_ms = retrieval_time, ) self.logger.info(f"Retrieved {len(chunks)} chunks in {retrieval_time:.0f}ms") # Determine prompt type and temperature self.logger.debug("Determining prompt strategy...") prompt_type = self._infer_prompt_type(query = request.query) temperature = self._get_adaptive_temperature(request = request, query = request.query, context = context, retrieval_scores = [chunk.score for chunk in chunks], prompt_type = prompt_type, ) self.logger.debug(f"Prompt type: {prompt_type.value}, Temperature: {temperature}") # Build optimized prompt self.logger.debug("Building prompt...") prompt = self.prompt_builder.build_prompt(query = request.query, context = context, sources = chunks, prompt_type = prompt_type, include_citations = request.include_sources, max_completion_tokens = request.max_tokens or self.settings.MAX_TOKENS, ) # Generate LLM response self.logger.debug("Generating LLM response...") generation_start = time.time() messages = [{"role" : "system", "content" : prompt["system"] }, {"role" : "user", "content" : prompt["user"], } ] llm_response = await self.llm_client.generate(messages = messages, temperature = temperature, top_p = request.top_p or self.settings.TOP_P, max_tokens = request.max_tokens or self.settings.MAX_TOKENS, ) generation_time = (time.time() - generation_start) * 1000 answer = llm_response["content"] self.logger.info(f"Generated response in {generation_time:.0f}ms ({llm_response['usage']['completion_tokens']} tokens)") # Format citations (if enabled) if request.include_sources: self.logger.debug("Formatting citations...") answer = self._post_process_citations(answer = answer, sources = chunks, ) # Create response object total_time = (time.time() - start_time) * 1000 response = QueryResponse(query = request.query, answer = answer, sources = chunks if request.include_sources else [], retrieval_time_ms = retrieval_time, generation_time_ms = generation_time, total_time_ms = total_time, tokens_used = {"input" : llm_response["usage"]["prompt_tokens"], "output" : llm_response["usage"]["completion_tokens"], "total" : llm_response["usage"]["total_tokens"], }, model_used = self.model_name, timestamp = datetime.now(), ) # Add quality metrics if requested if request.include_metrics: response.metrics = self._calculate_quality_metrics(query = request.query, answer = answer, context = context, sources = chunks, ) # Track both: prediction & reality response.metrics["predicted_type"] = classification.get('type', 'unknown') response.metrics["predicted_confidence"] = classification.get('confidence', 0.0) response.metrics["actual_type"] = "rag" # Always rag if we're here response.metrics["execution_path"] = "rag_pipeline" response.metrics["has_context"] = len(chunks) > 0 response.metrics["context_chunks"] = len(chunks) response.metrics["rag_confidence"] = min(1.0, sum(c.score for c in chunks) / len(chunks) if chunks else 0.0) response.metrics["is_forced_rag"] = classification.get('is_forced_rag', False) response.metrics["llm_classified"] = classification.get('is_llm_classified', False) # Add context for evaluation response.metrics["context_for_evaluation"] = context # Update statistics self.generation_count += 1 self.total_generation_time += total_time self.logger.info(f"RAG response generated successfully in {total_time:.0f}ms") return response except Exception as e: self.logger.error(f"RAG query handling failed: {repr(e)}", exc_info = True) if allow_fallback: # Return empty to trigger fallback return QueryResponse(query = request.query, answer = "", sources = [], retrieval_time_ms = 0.0, generation_time_ms = 0.0, total_time_ms = 0.0, tokens_used = {"input": 0, "output": 0, "total": 0}, model_used = self.model_name, timestamp = datetime.now(), ) else: raise @handle_errors(error_type = ResponseGenerationError, log_error = True, reraise = True) async def generate_response_stream(self, request: QueryRequest, has_documents: bool = True) -> AsyncGenerator[str, None]: """ Generate streaming RAG response Arguments: ---------- request { QueryRequest } : Query request object has_documents { bool } : Whether documents are available Yields: ------- { str } : Response chunks (tokens) """ self.logger.info(f"Generating streaming response for query: '{request.query[:100]}...'") try: # Classify query first classification = await self.query_classifier.classify(query = request.query, has_documents = has_documents, ) if (classification['suggested_action'] == 'respond_with_general_llm'): # Stream general response general_response = await self.general_responder.respond(query = request.query) yield general_response.get("answer", "") return # Otherwise proceed with RAG streaming - Procced with Retrieving context retrieval_result = self.hybrid_retriever.retrieve_with_context(query = request.query, top_k = request.top_k or self.settings.TOP_K_RETRIEVE, enable_reranking = request.enable_reranking, include_citations = request.include_sources, ) chunks = retrieval_result["chunks"] context = retrieval_result["context"] if not chunks: yield "I couldn't find relevant information to answer your question." return # Determine strategy prompt_type = self._infer_prompt_type(query = request.query) temperature = self._get_adaptive_temperature(request = request, query = request.query, context = context, retrieval_scores = [chunk.score for chunk in chunks], prompt_type = prompt_type, ) # Build prompt prompt = self.prompt_builder.build_prompt(query = request.query, context = context, sources = chunks, prompt_type = prompt_type, include_citations = request.include_sources, max_completion_tokens = request.max_tokens or self.settings.MAX_TOKENS, ) # Stream LLM response messages = [{"role" : "system", "content" : prompt["system"], }, {"role" : "user", "content" : prompt["user"], }, ] async for chunk_text in self.llm_client.generate_stream(messages = messages, temperature = temperature, top_p = request.top_p or self.settings.TOP_P, max_tokens = request.max_tokens or self.settings.MAX_TOKENS, ): yield chunk_text self.logger.info("Streaming response completed") except Exception as e: self.logger.error(f"Streaming generation failed: {repr(e)}", exc_info = True) yield f"\n\n[Error: {str(e)}]" def _infer_prompt_type(self, query: str) -> PromptType: """ Infer appropriate prompt type from query Arguments: ---------- query { str } : User query Returns: -------- { PromptType } : Inferred prompt type """ query_lower = query.lower() # Summary indicators if (any(word in query_lower for word in ['summarize', 'summary', 'overview', 'tldr', 'brief'])): return PromptType.SUMMARY # Comparison indicators if (any(word in query_lower for word in ['compare', 'contrast', 'difference', 'versus', 'vs'])): return PromptType.COMPARISON # Analytical indicators if (any(word in query_lower for word in ['analyze', 'analysis', 'evaluate', 'assess', 'examine'])): return PromptType.ANALYTICAL # Extraction indicators if (any(word in query_lower for word in ['extract', 'list', 'find all', 'identify', 'enumerate'])): return PromptType.EXTRACTION # Creative indicators if (any(word in query_lower for word in ['create', 'write', 'compose', 'generate', 'imagine'])): return PromptType.CREATIVE # Conversational indicators if (any(word in query_lower for word in ['tell me about', 'explain', 'discuss', 'talk about'])): return PromptType.CONVERSATIONAL # Default to QA return PromptType.QA def _get_adaptive_temperature(self, request: QueryRequest, query: str, context: str, retrieval_scores: List[float], prompt_type: PromptType) -> float: """ Get adaptive temperature based on query characteristics Arguments: ---------- request { QueryRequest } : Original request query { str } : User query context { str } : Retrieved context retrieval_scores { list } : Retrieval scores prompt_type { PromptType } : Inferred prompt type Returns: -------- { float } : Temperature value """ # Use request temperature if explicitly provided if (request.temperature is not None): self.logger.debug(f"Using request temperature: {request.temperature}") return request.temperature # Otherwise, use adaptive temperature controller temperature = self.temperature_controller.get_temperature(query = query, context = context, retrieval_scores = retrieval_scores, query_type = prompt_type.value, ) return temperature def _post_process_citations(self, answer: str, sources: List[ChunkWithScore]) -> str: """ Post-process answer to format citations Arguments: ---------- answer { str } : Generated answer with citation markers sources { list } : Source chunks Returns: -------- { str } : Answer with formatted citations """ try: # Validate citations is_valid, invalid = self.citation_formatter.validate_citations(answer, sources) if not is_valid: self.logger.warning(f"Invalid citations found: {invalid}") # Normalize to fix issues answer = self.citation_formatter.normalize_citations(answer, sources) # Format citations according to style formatted_answer = self.citation_formatter.format_citations_in_text(answer, sources) return formatted_answer except Exception as e: self.logger.error(f"Citation post-processing failed: {repr(e)}") # Return original answer if formatting fails return answer def _create_no_results_response(self, request: QueryRequest, retrieval_time_ms: float) -> QueryResponse: """ Create response when no results are found Arguments: ---------- request { QueryRequest } : Original request retrieval_time_ms { float } : Time spent on retrieval Returns: -------- { QueryResponse } : Response indicating no results """ no_results_answer = ("I couldn't find relevant information in the available documents to answer your question. " "This could mean:\n" "1. The information is not present in the indexed documents\n" "2. The question may need to be rephrased for better matching\n" "3. The relevant documents haven't been uploaded yet\n\n" "Please try:\n" "- Rephrasing your question with different keywords\n" "- Asking a more specific or general question\n" "- Ensuring the relevant documents are uploaded\n" ) return QueryResponse(query = request.query, answer = no_results_answer, sources = [], retrieval_time_ms = retrieval_time_ms, generation_time_ms = 0.0, total_time_ms = retrieval_time_ms, tokens_used = {"input": 0, "output": 0, "total": 0}, model_used = self.model_name, timestamp = datetime.now(), ) def _calculate_quality_metrics(self, query: str, answer: str, context: str, sources: List[ChunkWithScore]) -> Dict[str, float]: """ Calculate quality metrics for the response Arguments: ---------- query { str } : User query answer { str } : Generated answer context { str } : Retrieved context sources { list } : Source chunks Returns: -------- { dict } : Quality metrics """ metrics = dict() try: # Answer length metrics metrics["answer_length"] = len(answer.split()) metrics["answer_char_length"] = len(answer) # Citation metrics citation_stats = self.citation_formatter.get_citation_statistics(answer, sources) metrics["citations_used"] = citation_stats.get("total_citations", 0) metrics["unique_citations"] = citation_stats.get("unique_citations", 0) metrics["citation_density"] = citation_stats.get("citation_density", 0.0) # Context utilization context_length = len(context.split()) metrics["context_utilization"] = min(1.0, metrics["answer_length"] / max(1, context_length)) # Retrieval quality if sources: avg_score = sum(s.score for s in sources) / len(sources) metrics["avg_retrieval_score"] = avg_score metrics["top_retrieval_score"] = sources[0].score if sources else 0.0 # Query-answer alignment (simple keyword overlap) query_words = set(query.lower().split()) answer_words = set(answer.lower().split()) overlap = len(query_words & answer_words) metrics["query_answer_overlap"] = overlap / max(1, len(query_words)) except Exception as e: self.logger.warning(f"Failed to calculate some quality metrics: {repr(e)}") return metrics async def generate_batch_responses(self, requests: List[QueryRequest], has_documents: bool = True) -> List[QueryResponse]: """ Generate responses for multiple queries in batch Arguments: ---------- requests { list } : List of query requests has_documents { bool } : Whether documents are available Returns: -------- { list } : List of query responses """ self.logger.info(f"Generating batch responses for {len(requests)} queries") tasks = [self.generate_response(request = request, has_documents = has_documents) for request in requests] responses = await asyncio.gather(*tasks, return_exceptions = True) # Handle exceptions results = list() for i, response in enumerate(responses): if isinstance(response, Exception): self.logger.error(f"Batch query {i} failed: {repr(response)}") # Create error response error_response = self._create_error_response(requests[i], str(response)) results.append(error_response) else: results.append(response) self.logger.info(f"Completed batch generation: {len(results)} responses") return results def _create_error_response(self, request: QueryRequest, error_message: str) -> QueryResponse: """ Create error response for failed generation Arguments: ---------- request { QueryRequest } : Original request error_message { str } : Error message Returns: -------- { QueryResponse } : Error response """ return QueryResponse(query = request.query, answer = f"An error occurred while generating the response: {error_message}", sources = [], retrieval_time_ms = 0.0, generation_time_ms = 0.0, total_time_ms = 0.0, tokens_used = {"input": 0, "output": 0, "total": 0}, model_used = self.model_name, timestamp = datetime.now(), ) def get_generation_stats(self) -> Dict: """ Get generation statistics including query type breakdown Returns: -------- { dict } : Generation statistics """ avg_time = (self.total_generation_time / self.generation_count) if self.generation_count > 0 else 0 return {"total_generations" : self.generation_count, "general_queries" : self.general_query_count, "rag_queries" : self.rag_query_count, "total_generation_time" : self.total_generation_time, "avg_generation_time_ms" : avg_time, "provider" : self.provider.value, "model" : self.model_name, "llm_health" : self.llm_client.check_health(), "query_routing_enabled" : True, "llm_based_routing" : True, } def reset_stats(self): """ Reset generation statistics """ self.generation_count = 0 self.general_query_count = 0 self.rag_query_count = 0 self.total_generation_time = 0.0 self.logger.info("Generation statistics reset") # Global response generator instance _response_generator = None def get_response_generator(provider: LLMProvider = None, model_name: str = None) -> ResponseGenerator: """ Get global response generator instance (singleton) Arguments: ---------- provider { LLMProvider } : LLM provider model_name { str } : Model name Returns: -------- { ResponseGenerator } : ResponseGenerator instance """ global _response_generator if _response_generator is None or (provider and _response_generator.provider != provider): _response_generator = ResponseGenerator(provider, model_name) return _response_generator @handle_errors(error_type = ResponseGenerationError, log_error = True, reraise = False) async def generate_answer(query: str, top_k: int = 5, temperature: float = None, has_documents: bool = True, **kwargs) -> str: """ Convenience function for quick answer generation Arguments: ---------- query { str } : User query top_k { int } : Number of chunks to retrieve temperature { float } : Temperature for generation has_documents { bool } : Whether documents are available **kwargs : Additional parameters Returns: -------- { str } : Generated answer """ request = QueryRequest(query = query, top_k = top_k, temperature = temperature, **kwargs ) generator = get_response_generator() response = await generator.generate_response(request = request, has_documents = has_documents, ) return response.answer