Multi-Version Document Q&A with RAG

A RAG system with built-in version control enabling semantic search and natural language Q&A across document revision history.

Overview

Solves document version tracking by implementing version-aware semantic search. Unlike traditional RAG systems that only support current versions, this system maintains complete document history with temporal querying capabilities.

Key Features

• Version Control: Query any historical version ("What did v3 say about remote work?")
• Temporal Comparison: Compare any two versions with automated change detection
• Incremental Indexing: O(1) version additions without index rebuilds
• LLM Integration: Natural language answers via Llama 3.3-70B (Groq API)
• Change Detection: Automatic classification of modified/added/removed content

---

Tech Stack

• Backend: FastAPI, FAISS, Sentence Transformers, SQLAlchemy, Groq API
• Frontend: React
• Deployment: Hugging Face Spaces (backend API), Vercel (frontend)
• Containerization: Docker

---

Architecture

React Frontend (Vercel)
    ↓ HTTPS
FastAPI Backend (Hugging Face Spaces / Docker)
    ↓
[SQLite] [FAISS Index] [Groq LLM]

Core Components

1. Document Processor: Extracts text, chunks (512 chars, 50 overlap)
2. Embeddings: Sentence Transformers (all-MiniLM-L6-v2, 384 dims)
3. Vector Store: FAISS with incremental indexing
4. Version Metadata: SQLAlchemy ORM (Document, Version, Chunk tables)
5. LLM Integration: Groq API for answer generation

---

Installation

Backend Setup (Local)

cd server
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

export GROQ_API_KEY=YOUR_GROQ_API_KEY

python server_app.py

Runs at http://localhost:8000

Frontend Setup

cd ui
npm install
npm start

Runs at http://localhost:3000

---

API Endpoints

Upload Document

POST /api/documents/upload
Content-Type: multipart/form-data

file: [PDF/TXT/DOCX]
doc_name: "policy"

Query

POST /api/query/generate
{
  "question": "What is the remote work policy?",
  "version_id": 2,
  "k": 5
}

Compare Versions

POST /api/compare/detailed
{
  "question": "What changed?",
  "version_id_1": 1,
  "version_id_2": 2
}

List Versions

GET /api/documents/{doc_name}/versions

---

How It Works

1. Document Upload

• Extract text (PyPDF, python-docx)
• Chunk text (512 chars, 50 overlap)
• Generate embeddings (Sentence Transformers)
• Store in FAISS (incremental add) + SQLite (metadata)

2. Query

• Generate question embedding
• FAISS similarity search (filter by version_id)
• LLM generates answer from retrieved chunks

3. Version Comparison

• Query both versions
• Calculate chunk-level cosine similarity
• Classify changes (modified/added/removed)
• LLM summarizes differences

---

Project Structure

├── server/
│   ├── src/
│   │   ├── database.py           # SQLAlchemy models
│   │   ├── embeddings.py         # Sentence Transformers wrapper
│   │   ├── vector_store.py       # FAISS operations
│   │   ├── document_processor.py # Text extraction & chunking
│   │   └── rag_system.py         # Main orchestrator
│   ├── server_app.py             # FastAPI application
│   └── requirements.txt
│
└── ui/
    ├── public/
    ├── src/
    │   └── components/
    │   ├── App.js
    │   ├── config.js
    │   ├── index.js
    └── package.json

---

Deployment

Hugging Face Spaces (Backend)

This project is deployed on Hugging Face Spaces using the Docker Space SDK on the free CPU tier.

Space URL: https://technophyle-doc-qa.hf.space/

Deployment details

• Space SDK: Docker
• Runtime: FastAPI (Uvicorn)
• Containerization: Docker
• Hardware: Free CPU

Required environment variables: GROQ_API_KEY

Vercel (Frontend)

Framework: React

---

Use Cases

• Policy Tracking: Track changes in HR policies, contracts, regulations
• Compliance: Audit document revisions with searchable history
• Legal: Compare contract versions, track amendments
• Knowledge Management: Maintain evolving documentation with temporal queries

---