# 🏛️ PRODUCTION ENTERPRISE RAG ARCHITECTURE BLUEPRINT ### PHASE 1: FORENSIC SOURCE AUDIT & INGESTION PIPELINE * Source Classification: - Authoritative: HR Policies (PDF/DOCX), Employee Handbook. - Derived/Secondary: Internal HR FAQ Database. - Restricted Authoritative: Personal Compensation & Employee Records. * Ingestion Guardrail: Quarantine queue isolates malformed PDFs; layout-parser preserves heading hierarchies (H1->H3) and tabular cells (Expense Limit vs Condition). ### PHASE 2: STRUCTURE-AWARE CHUNKING & METADATA PROVENANCE * Strategy: Hierarchical Parent-Child Chunking. - Parent: Complete Policy Clause (Context Anchor). - Child: Atomic Rule/Condition (Retrieval Unit). * Metadata Schema: `{"doc_id": "POL-LEAVE-2026", "version": "3.1", "effective_date": "2026-01-01", "role_access": ["all_employees"], "section": "Casual Leave > Eligibility"}` ### PHASE 3: HYBRID INDEXING & SECURE RETRIEVAL PIPELINE * Dual Index: Qdrant Vector (OpenAI text-embedding-3-large) + BM25 Lexical Engine. * Security Filter: Pre-retrieval filtering restricts search queries to chunks where `role_access` matches user JWT claim. * Reranker: Cohere Rerank v3 filters top-25 hybrid candidates down to top-5 verified evidence passages. ### PHASE 4: AGENTIC REASONING & CONFLICT RESOLUTION * Conflict Engine: If version mismatch is detected between Handbook and FAQ, the document with the latest `effective_date` and `Authoritative` tag explicitly invalidates secondary FAQ entries. * Hallucination Gate: If no retrieved chunk matches user eligibility criteria with cosine similarity > 0.72, return: "The current HR policies do not establish this entitlement." ### PHASE 5: TARGET ARCHITECTURE & PYTHON SKELETON User Query ──► [Auth Context Extractor] │ (user_id, role) ▼ [Pre-Filtered Hybrid Search] ┌──────────┴──────────┐ ▼ ▼ [BM25 Engine] [Vector DB (Qdrant)] └──────────┬──────────┘ ▼ [Cross-Encoder Reranker] │ (Top-5 chunks) ▼ [Grounding & Citation Validator] ──► Final Answer + Section Footnotes ```python from pydantic import BaseModel, Field from typing import List, Optional class ChunkMetadata(BaseModel): doc_id: str version: str effective_date: str role_access: List[str] section_ref: str class GroundedCitation(BaseModel): statement: str source_doc: str section: str version: str class RAGResponse(BaseModel): answer: str citations: List[GroundedCitation] is_sufficient_evidence: bool