from __future__ import annotations import logging import os import re from typing import Any, Dict, List, Optional from ..query_metadata_extraction import QueryMetadata, QueryMetadataExtractor logger = logging.getLogger(__name__) # Phrases that indicate the user is explicitly requesting date-scoped documents _DATE_FILTER_PHRASES = ( "documents from", "papers from", "published in", "dated", "in the year", "from the year", "filed in", "issued in", "show me", "find me", "retrieve", "documents dated", "from 20", "from 19", ) class AdaptiveQueryMetadataExtractor: """ Extracts explicit metadata constraints from user queries. Uses open source models via Groq or Hugging Face, with deterministic regex-based extraction as a fallback. Each extracted field carries a '_confidence' sub-key indicating whether the field was explicitly requested ('high'), heuristically inferred ('medium'), or weakly implied ('low'). The QueryPlanner uses this to decide whether the field becomes a hard filter or a soft ranking hint. """ def __init__(self, use_llm_if_available: bool = True) -> None: self.use_llm = use_llm_if_available self._groq_extractor: Optional[Any] = None self._hf_extractor: Optional[QueryMetadataExtractor] = None if self.use_llm and os.getenv("GROQ_API_KEY") and not os.getenv("OFFLINE_EVAL"): try: from langchain_groq import ChatGroq model_name = os.getenv("GROQ_MODEL", "qwen/qwen3.8-27b") self._groq_extractor = ChatGroq( model=model_name, api_key=os.getenv("GROQ_API_KEY"), temperature=0.0, max_retries=1, request_timeout=5, ).with_structured_output(QueryMetadata) except Exception as exc: logger.warning(f"Could not initialize Groq metadata extractor: {exc}") if self.use_llm and os.getenv("HF_TOKEN") and not self._groq_extractor and not os.getenv("OFFLINE_EVAL"): try: self._hf_extractor = QueryMetadataExtractor() except Exception as exc: logger.warning(f"Could not initialize HF metadata extractor: {exc}") def extract(self, query: str) -> Dict[str, Any]: # Fast sub-millisecond heuristic check first rule_extracted = self._extract_rule_based(query) if ( rule_extracted["document_type"] or rule_extracted["organizations"] or rule_extracted["dates"] or os.getenv("OFFLINE_EVAL") ): return rule_extracted if self._groq_extractor: try: import json from pathlib import Path prompt_file = Path(__file__).resolve().parents[3] / "prompts" / "query_metadata_extraction.json" system_prompt = "Extract explicit metadata constraints from the query. Return only JSON matching schema." if prompt_file.exists(): data = json.loads(prompt_file.read_text(encoding="utf-8")) system_prompt = data.get("query_metadata_extraction", {}).get("system", system_prompt) extracted = self._groq_extractor.invoke([ ("system", system_prompt), ("human", query), ]) # LLM-extracted fields are HIGH confidence (user explicitly stated) result = { "document_type": extracted.document_type or "", "organizations": extracted.organizations or [], "locations": extracted.locations or [], "dates": extracted.dates or [], "department": extracted.department or "", "topics": extracted.topics or [], "_confidence": { "document_type": "high", "organizations": "high", "dates": "high", "locations": "high", "department": "high", }, } return result except Exception as exc: logger.warning(f"Groq metadata extraction failed: {exc}, disabling LLM extractor.") self._groq_extractor = None if self._hf_extractor: try: extracted = self._hf_extractor.extract(query) result = { "document_type": extracted.document_type, "organizations": extracted.organizations, "locations": extracted.locations, "dates": extracted.dates, "department": extracted.department, "topics": extracted.topics, "_confidence": { "document_type": "high", "organizations": "high", "dates": "high", "locations": "high", "department": "high", }, } return result except Exception as exc: logger.warning(f"HF metadata extraction failed: {exc}, using rule-based.") return self._extract_rule_based(query) def _extract_rule_based(self, query: str) -> Dict[str, Any]: """ Conservative rule-based extraction. Each extracted field is annotated with a confidence level: "high" - the user explicitly requested a document-scoped filter "medium" - likely a filter but not certain "low" - heuristic mention; should NOT be applied as a hard filter The QueryPlanner is responsible for honouring only HIGH-confidence fields as hard Qdrant/BM25 filters. """ q = query.lower() confidence: Dict[str, str] = {} extracted: Dict[str, Any] = { "document_type": "", "organizations": [], "locations": [], "dates": [], "department": "", "topics": [], } # ── Document type ────────────────────────────────────────────────── # Explicit document-type requests ("find resumes", "show papers", …) if "resume" in q or "cv" in q: extracted["document_type"] = "resume" confidence["document_type"] = "high" elif "application form" in q or "application instruction" in q: extracted["document_type"] = "application instructions" confidence["document_type"] = "high" elif "research paper" in q or "nips" in q: extracted["document_type"] = "research_paper" confidence["document_type"] = "high" elif ("paper" in q or "publication" in q) and any( kw in q for kw in ("show", "find", "retrieve", "published", "authored") ): extracted["document_type"] = "research_paper" confidence["document_type"] = "medium" elif "paper" in q: # "paper" alone inside a content question is NOT a type filter # e.g. "What did the paper say about attention?" - keep as topic only confidence["document_type"] = "low" elif ("report" in q or "financial report" in q) and any( kw in q for kw in ("show", "find", "retrieve") ): extracted["document_type"] = "report" confidence["document_type"] = "medium" # ── Years / dates ────────────────────────────────────────────────── years = re.findall(r"\b(19\d{2}|20\d{2})\b", query) if years: extracted["dates"] = list(set(years)) # Apply as hard filter only when the user explicitly scopes by date if any(ph in q for ph in _DATE_FILTER_PHRASES): confidence["dates"] = "high" else: # Year appears as context (e.g. "the 2017 Transformer paper") # - treat as soft retrieval hint only confidence["dates"] = "low" # ── Organizations ────────────────────────────────────────────────── if "niti aayog" in q: extracted["organizations"].append("NITI Aayog") confidence["organizations"] = "high" elif "niti" in q and any( kw in q for kw in ("internship", "scheme", "guidelines", "policy", "aayog") ): extracted["organizations"].append("NITI Aayog") confidence["organizations"] = "medium" elif "niti" in q: extracted["organizations"].append("NITI Aayog") confidence["organizations"] = "low" if "google" in q: extracted["organizations"].append("Google") # Mentioning "Google researchers" is about paper authorship, not a doc filter existing = confidence.get("organizations", "low") # Don't upgrade confidence if already set by NITI if "organizations" not in confidence: confidence["organizations"] = "low" # ── Topics (soft signals only - never hard filters) ──────────────── match = re.search(r"\b(?:about|regarding|on)\s+([a-zA-Z0-9\s\-]+)", query, re.IGNORECASE) if match: topic = match.group(1).strip() if topic: extracted["topics"].append(topic) extracted["_confidence"] = confidence return extracted