"""Generation and Response Module for Document Question Answering. Features: - Strict evidence grounding (answers ONLY from retrieved chunks) - Quotes evidence and cites exact [Page X, Section "Y"] - If answer is not in context, outputs "Not found in the document." - Streaming support for responsive chat interfaces - Conversation memory tracking - Integrated second-pass verification - Multi-backend LLM support: 1. Hugging Face Serverless Inference API (Qwen/Qwen2.5-7B, Llama-3.2, Mistral) 2. OpenAI / OpenAI-compatible APIs (OpenAI, Groq, Ollama, DeepSeek) 3. Offline Deterministic Grounding Engine (0 external calls, 100% cited) """ import json import os import re import time from dataclasses import dataclass, field from typing import List, Dict, Any, Optional, Iterator, Tuple from ingest import DocumentChunk from verify import AnswerVerifier, VerificationResult SYSTEM_PROMPT = """You are a production-grade Document Question Answering Assistant. CRITICAL INSTRUCTIONS: 1. Answer the user's question ONLY using the factual information in the PROVIDED CONTEXT below. 2. Do NOT extrapolate, assume, or use any prior or external knowledge outside the provided context. 3. If the answer cannot be fully determined from the provided context, respond with EXACTLY: "Not found in the document." 4. Always quote direct evidence or key statistics from the context when formulating your answer. 5. For EVERY factual statement you make, append an exact citation in the format: [Page , Section ""] 6. Maintain a professional, concise, and structured tone. """ def format_context_for_prompt(chunks: List[DocumentChunk]) -> str: """Formats retrieved chunks with provenance metadata into an LLM prompt block.""" if not chunks: return "No relevant context retrieved." blocks = [] for idx, chunk in enumerate(chunks, 1): block = ( f"--- Context Chunk {idx} ---\n" f"Document: {chunk.file_name}\n" f"Page: {chunk.page_number}\n" f"Section: {chunk.section_title}\n" f"Text:\n{chunk.text.strip()}\n" ) blocks.append(block) return "\n".join(blocks) def normalize_word(w: str) -> str: """Lightweight rule-based stemmer to match plurals, past tense, and suffixes.""" w = w.lower().strip() if len(w) > 4: if w.endswith("ies"): return w[:-3] + "y" if w.endswith("es"): return w[:-2] if w.endswith("s") and not w.endswith("ss"): return w[:-1] if w.endswith("ing"): return w[:-3] if w.endswith("ed"): return w[:-2] return w class OfflineGroundingLLM: """ Offline deterministic grounding engine. Ensures the QA pipeline runs out of the box with zero external API dependencies, enforcing 100% strict adherence to context and citations. """ def generate(self, prompt: str, context_chunks: List[DocumentChunk]) -> str: if not context_chunks: return "Not found in the document." # Extract user query from prompt match = re.search(r"User Question:\s*(.+?)(?:\n|$)", prompt) query = match.group(1).strip() if match else prompt.strip() query_words = set(re.findall(r"\b[a-zA-Z0-9_]{3,}\b", query.lower())) # Remove common question words and prepositions stop_words = { "what", "when", "where", "which", "who", "whom", "whose", "why", "how", "is", "are", "was", "were", "be", "been", "being", "have", "has", "had", "do", "does", "did", "the", "a", "an", "and", "or", "in", "on", "at", "to", "for", "of", "with", "by", "from", "can", "could", "should", "would", "please", "about", "tell", "explain", "describe", "find", "show", "many", "much" } query_keywords = query_words - stop_words if not query_keywords: return "Not found in the document." query_stems = {normalize_word(w) for w in query_keywords} # Document filename generic words (e.g. "Hyperledger", "Notes", "report", "cloud") doc_generic_terms = set() for c in context_chunks: base = os.path.splitext(c.file_name)[0] for w in re.findall(r"\b[a-zA-Z]{3,}\b", base.lower()): doc_generic_terms.add(normalize_word(w)) core_stems = query_stems - doc_generic_terms if not core_stems: core_stems = set(query_stems) # Helper to detect textbook definitional statements for specific core concepts def is_definitional_sentence(sentence: str) -> bool: s_lower = sentence.lower().strip() target_keywords = (query_keywords - doc_generic_terms) if len(query_keywords) >= 2 else query_keywords for kw in target_keywords: kw_clean = kw.lower().rstrip("s") pattern = rf"\b{kw_clean}s?\s+(?:is|are|refers to|means|defines)\b" if re.search(pattern, s_lower): return True return False # Scan chunks for matching information matching_sentences: List[Tuple[str, DocumentChunk, float]] = [] for chunk_idx, chunk in enumerate(context_chunks): rank_multiplier = 1.0 / (1.0 + chunk_idx * 0.15) # Check for table chunk if chunk.metadata.get("is_table", False): lines = chunk.text.splitlines() for line in lines: line_words = set(re.findall(r"\b[a-zA-Z0-9_]{3,}\b", line.lower())) line_stems = {normalize_word(w) for w in line_words} overlap = len(query_stems & line_stems) exact_overlap = len(query_keywords & line_words) core_overlap = len(core_stems & line_stems) if len(core_stems) > 0 and core_overlap == 0 and len(query_keywords) >= 2: continue if overlap >= 1: score = ((overlap ** 2) * 3.0 + (core_overlap ** 2) * 5.0 + exact_overlap * 2.0) * rank_multiplier matching_sentences.append((line.strip(), chunk, score)) continue # Split on sentence enders AND newlines (handling notes, bullet points, slides) sentences = re.split(r"(?<=[.!?])\s+|\n+", chunk.text) for s in sentences: s_clean = re.sub(r"^[\s*\-•#\d+.)]+", "", s).strip() if len(s_clean.split()) < 3: continue # Filter out running page headers / course boilerplate if re.search(r"\b(B\.Tech|ICFAI|Tech School|Lecture Notes|All rights reserved)\b", s_clean, re.IGNORECASE): continue s_words = set(re.findall(r"\b[a-zA-Z0-9_]{3,}\b", s_clean.lower())) s_stems = {normalize_word(w) for w in s_words} overlap = len(query_stems & s_stems) exact_overlap = len(query_keywords & s_words) core_overlap = len(core_stems & s_stems) # Disqualify sentences that match ZERO specific core keywords when specific keywords exist if len(core_stems) > 0 and core_overlap == 0 and len(query_keywords) >= 2: continue if overlap >= 1: score = ((overlap ** 2) * 4.0 + (core_overlap ** 2) * 6.0 + exact_overlap * 2.0) * rank_multiplier # Definitional boost for core query concepts if is_definitional_sentence(s_clean): score += 20.0 matching_sentences.append((s_clean, chunk, score)) if not matching_sentences: return "Not found in the document." # Sort sentences by match score descending matching_sentences.sort(key=lambda x: x[2], reverse=True) # Deduplicate sentences while preserving citation seen = set() answer_parts = [] for s_text, chunk, _ in matching_sentences: cleaned_key = re.sub(r"[^a-z0-9]", "", s_text.lower()) if cleaned_key in seen: continue seen.add(cleaned_key) citation = f'[Page {chunk.page_number}, Section "{chunk.section_title}"]' answer_parts.append(f"{s_text} {citation}") if len(answer_parts) >= 3: break if not answer_parts: return "Not found in the document." return " ".join(answer_parts) class HuggingFaceInferenceClient: """Client for Hugging Face Serverless Inference API.""" def __init__( self, model: str = "Qwen/Qwen2.5-7B-Instruct", token: Optional[str] = None, temperature: float = 0.01, max_tokens: int = 1024 ): self.model = model self.token = token or os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACEHUB_API_TOKEN") self.temperature = max(0.01, temperature) self.max_tokens = max_tokens self._client = None try: from huggingface_hub import InferenceClient self._client = InferenceClient(model=self.model, token=self.token) except Exception: self._client = None def is_configured(self) -> bool: return self._client is not None def stream_chat(self, messages: List[Dict[str, str]]) -> Iterator[str]: if not self._client: raise RuntimeError("Hugging Face client is not configured.") response = self._client.chat_completion( messages=messages, temperature=self.temperature, max_tokens=self.max_tokens, stream=True ) for chunk in response: content = chunk.choices[0].delta.content or "" if content: yield content def complete_chat(self, messages: List[Dict[str, str]]) -> str: if not self._client: raise RuntimeError("Hugging Face client is not configured.") response = self._client.chat_completion( messages=messages, temperature=self.temperature, max_tokens=self.max_tokens, stream=False ) return response.choices[0].message.content or "" class OpenAILLMClient: """Wrapper for OpenAI or OpenAI-compatible APIs (Groq, Ollama, DeepSeek, vLLM).""" def __init__( self, api_key: Optional[str] = None, base_url: Optional[str] = None, model: str = "gpt-4o-mini", temperature: float = 0.0, max_tokens: int = 1024 ): self.api_key = api_key or os.getenv("OPENAI_API_KEY") self.base_url = base_url or os.getenv("OPENAI_BASE_URL") self.model = model self.temperature = temperature self.max_tokens = max_tokens self._client = None if self.api_key: try: import openai self._client = openai.OpenAI( api_key=self.api_key, base_url=self.base_url ) except Exception: self._client = None def is_configured(self) -> bool: return self._client is not None def stream_chat(self, messages: List[Dict[str, str]]) -> Iterator[str]: if not self._client: raise RuntimeError("OpenAI client is not configured.") response = self._client.chat.completions.create( model=self.model, messages=messages, temperature=self.temperature, max_tokens=self.max_tokens, stream=True ) for chunk in response: content = chunk.choices[0].delta.content or "" if content: yield content def complete_chat(self, messages: List[Dict[str, str]]) -> str: if not self._client: raise RuntimeError("OpenAI client is not configured.") response = self._client.chat.completions.create( model=self.model, messages=messages, temperature=self.temperature, max_tokens=self.max_tokens, stream=False ) return response.choices[0].message.content or "" class ConversationMemory: """Manages chat history and tracks citations per conversation turn.""" def __init__(self, max_turns: int = 10): self.max_turns = max_turns self.history: List[Dict[str, Any]] = [] def add_user_message(self, message: str) -> None: self.history.append({"role": "user", "content": message.strip()}) self._trim() def add_assistant_message( self, message: str, citations: Optional[List[Dict[str, Any]]] = None, verification: Optional[VerificationResult] = None ) -> None: turn = { "role": "assistant", "content": message.strip(), "citations": citations or [], "verification": verification.to_dict() if verification else None, } self.history.append(turn) self._trim() def get_messages_for_prompt(self) -> List[Dict[str, str]]: """Returns message list formatted for chat completions.""" return [{"role": m["role"], "content": m["content"]} for m in self.history] def _trim(self) -> None: if len(self.history) > self.max_turns * 2: self.history = self.history[- (self.max_turns * 2):] def clear(self) -> None: self.history.clear() class Generator: """ Main Document QA Generator with verification and streaming support. """ def __init__( self, llm_client: Optional[Any] = None, verifier: Optional[AnswerVerifier] = None, enable_verification: bool = True ): # Auto-select client if none provided if llm_client is None: if os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACEHUB_API_TOKEN"): self.llm_client = HuggingFaceInferenceClient() elif os.getenv("OPENAI_API_KEY"): self.llm_client = OpenAILLMClient() else: self.llm_client = None else: self.llm_client = llm_client self.offline_engine = OfflineGroundingLLM() self.verifier = verifier or AnswerVerifier() self.enable_verification = enable_verification self.memory = ConversationMemory() def build_prompt_messages(self, query: str, context_chunks: List[DocumentChunk]) -> List[Dict[str, str]]: """Constructs system and user messages incorporating retrieved context.""" context_str = format_context_for_prompt(context_chunks) user_content = ( f"PROVIDED CONTEXT:\n{context_str}\n\n" f"User Question: {query}\n\n" "Answer using ONLY the provided context above. Quote direct evidence and cite [Page X, Section Y] for every statement." ) messages = [{"role": "system", "content": SYSTEM_PROMPT}] for turn in self.memory.history[-4:]: messages.append({"role": turn["role"], "content": turn["content"]}) messages.append({"role": "user", "content": user_content}) return messages def generate_answer( self, query: str, context_chunks: List[DocumentChunk], verify: bool = True ) -> Tuple[str, Optional[VerificationResult]]: """Synchronously generates a verified, grounded answer.""" messages = self.build_prompt_messages(query, context_chunks) # 1. Attempt LLM generation answer = "" if self.llm_client is not None and hasattr(self.llm_client, "is_configured") and self.llm_client.is_configured(): try: answer = self.llm_client.complete_chat(messages) except Exception: answer = "" # 2. Fallback to offline deterministic grounding engine if not answer: answer = self.offline_engine.generate( prompt=f"User Question: {query}", context_chunks=context_chunks ) # 3. Second-pass Verification verification = None if verify and self.enable_verification: verification = self.verifier.verify_answer(answer, context_chunks) answer = verification.verified_answer return answer, verification def stream_answer( self, query: str, context_chunks: List[DocumentChunk], chunk_delay_sec: float = 0.02 ) -> Iterator[str]: """Streams answer token by token for UI interactions.""" messages = self.build_prompt_messages(query, context_chunks) # If LLM configured, stream from client if self.llm_client is not None and hasattr(self.llm_client, "is_configured") and self.llm_client.is_configured(): try: for token in self.llm_client.stream_chat(messages): yield token return except Exception: pass # Offline streaming emulation answer = self.offline_engine.generate( prompt=f"User Question: {query}", context_chunks=context_chunks ) words = answer.split(" ") for i, word in enumerate(words): yield word + (" " if i < len(words) - 1 else "") time.sleep(chunk_delay_sec)