RAG Document Q&A

Ask questions about your PDF documents in plain English. Upload any document and get accurate, cited answers — powered by semantic search and Google Gemini.

Built to explore production patterns in LLM systems: vector retrieval, streaming responses, session isolation, and grounded generation.

---

Architecture

UPLOAD FLOW
─────────────────────────────────────────────────────────────
PDF ──► PyMuPDF ──► Chunks (512 tok, 50 overlap) ──► Gemini Embeddings ──► ChromaDB
                                                                               │
                                                                    (keyed by session ID)

QUERY FLOW
─────────────────────────────────────────────────────────────
User Question ──► Gemini Embeddings ──► Cosine Similarity Search ──► Top 5 Chunks
                                                                           │
                                                              ┌────────────▼────────────┐
                                                              │      Gemini LLM          │
                                                              │  + System Prompt         │
                                                              │  + Conversation History  │
                                                              │    (last 5 messages)     │
                                                              └────────────┬────────────┘
                                                                           │
                                                              Streamed Answer + Citations

---

Features

• Semantic Search — vector similarity retrieval via ChromaDB, not keyword matching
• Streaming Responses — token-by-token generation via Server-Sent Events
• Source Citations — every answer links back to the exact page and chunk it came from
• Conversation History — sliding window of last 5 messages for context-aware follow-ups
• Multi-User Sessions — UUID-based session isolation; each user's documents are namespaced separately
• Grounded Generation — system prompt enforces answers only from retrieved context, eliminating hallucination
• Production Patterns — rate limiting, structured logging, retry logic with exponential backoff

---

Tech Stack

Layer	Technology
Backend	Python, FastAPI
Vector Store	ChromaDB
LLM + Embeddings	Google Gemini API
PDF Parsing	PyMuPDF
Token Counting	tiktoken
Frontend	React, Tailwind CSS, Vite
Containerization	Docker, Docker Compose

---

Getting Started

Prerequisites

• Docker and Docker Compose
• Google Gemini API key (get one here)

Run with Docker

git clone https://github.com/kshitizj03/rag-pdf-chat
cd RAG

cp .env.example .env
# Add your GEMINI_API_KEY to .env

docker compose up --build

• Frontend: http://localhost:3000
• Backend API: http://localhost:8000

Local Development

Backend

cd backend
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
uvicorn main:app --reload --port 8000

Frontend

cd frontend
npm install
cp .env.example .env
npm run dev

---

License

MIT