Human-in-the-Loop Thematic RAG: Transparent Retrieval for Canonical Texts

Human-in-the-Loop Thematic RAG: Transparent Retrieval for Canonical Texts

The Thematic RAG System is an intelligent, human-in-the-loop retrieval and synthesis platform engineered for high-precision exploration of dense canonical, sacred, and regulatory texts. By decomposing question-answering into an explicit, multi-stage stateful graph, the system eliminates hallucinations and bridges the linguistic divide between modern queries and classical knowledge bases.

Problem Solved

Standard Retrieval-Augmented Generation (RAG) architectures fail when applied to structured, high-stakes literature:

• Arbitrary Document Chunking: Conventional pipelines slice text by fixed character or token counts, slicing across sentences, chapters, and canonical verse boundaries. This destroys contiguous context and produces disconnected fragments.
• The Semantic & Linguistic Gap: Users typically formulate questions in colloquial, modern phrasing, whereas target knowledge bases use classical terminology, specialized jurisprudence, and distinct morphological roots. Direct embedding searches fail to capture conceptual nuance.
• The Black-Box Hallucination Trap: Generic AI chatbots take a prompt, execute a blind vector search, and immediately generate an answer. When context is ambiguous, the LLM hallucinates citations, misattributes quotes, and produces plausible-sounding falsehoods.

Impact & Engineering Benchmarks

Engineered for rigorous evidence verification, low-latency search, and transparent user agency:

• Strict Contextual Grounding: Constrained the generation engine strictly to user-approved, retrieved excerpts with explicit instructions to refuse synthesis when contextual evidence is lacking.
• Sub-Second Vector Search & Context Hydration: Deployed dense embedding search in Pinecone paired with sub-50ms relational database lookups to rehydrate entire contiguous passages surrounding matched coordinates.
• Transparent Step-by-Step Inspection: Exposed intermediate pipeline steps (thematic expansions, extracted keywords, retrieved textual coordinates) directly to the user interface, enabling real-time review and manual refinement prior to generation.
• Production Cloud Scalability: Containerized the FastAPI orchestration service and deployed it via AWS Copilot onto auto-scaling AWS ECS Fargate clusters backed by CloudFormation infrastructure.

Solution

We architected a stepped, stateful LangGraph workflow that separates query comprehension from synthesis:

• Decomposed 4-Step Graph Workflow:
  1. Thematic Query Expansion: Translates colloquial user prompts into theological and conceptual domain themes while strictly prohibiting premature answers.
  2. Keyword & Scope Extraction: Analyzes expanded concepts to extract structured entity targets and canonical vocabulary.
  3. Structural Canonical Retrieval: Executes vector similarity search in Pinecone and rehydrates contiguous surrounding context using structured database coordinates (chapter_id, verse_start, verse_end).
  4. Constrained Grounded Generation: Feeds verified, user-inspected passages into the LLM with strict grounding prompts that eliminate confabulation.
• Modern Inspectable UX: Built a responsive Next.js frontend with Firebase Authentication, step-by-step progress steppers, inline editable keyword chips, and interactive source document selection cards.
• Cross-Lingual Concordance Roadmap: Architected morphological lookup tables enabling lexical search and root-word derivations across languages directly alongside semantic vector search.

Tech Stack

• AI & Orchestration: Python, LangChain, LangGraph, OpenAI / HuggingFace Embeddings
• Vector & Storage: Pinecone (Persistent Vector Database), SQLite / PostgreSQL (Relational Metadata)
• API & Frontend: FastAPI, Next.js, React, Tailwind CSS, Firebase Authentication
• Infrastructure & DevOps: AWS ECS (Fargate), AWS Copilot, Docker, AWS CloudFormation