Skip to content

README.md

Latest commit

 

History

History
164 lines (120 loc) · 6.49 KB

README.md

File metadata and controls

164 lines (120 loc) · 6.49 KB
Raw
██╗  ██╗██╗   ██╗██████╗ ██████╗ ██╗██████╗     ██████╗  █████╗  ██████╗
██║  ██║╚██╗ ██╔╝██╔══██╗██╔══██╗██║██╔══██╗    ██╔══██╗██╔══██╗██╔════╝
███████║ ╚████╔╝ ██████╔╝██████╔╝██║██║  ██║    ██████╔╝███████║██║  ███╗
██╔══██║  ╚██╔╝  ██╔══██╗██╔══██╗██║██║  ██║    ██╔══██╗██╔══██║██║   ██║
██║  ██║   ██║   ██████╔╝██║  ██║██║██████╔╝    ██║  ██║██║  ██║╚██████╔╝
╚═╝  ╚═╝   ╚═╝   ╚═════╝ ╚═╝  ╚═╝╚═╝╚═════╝     ╚═╝  ╚═╝╚═╝  ╚═╝ ╚═════╝

Ask anything. Get answers from your own documents.

Live Demo Python Streamlit Groq

⚡ What is this?

A production-ready Hybrid RAG system that combines the best of two retrieval worlds — semantic understanding + keyword precision — to answer questions from your documents with pinpoint accuracy.

No hallucinations. No generic answers. Just your data, intelligently retrieved.

Your PDF  ──▶  Dense Search (meaning)  ──┐
                                          ├──▶  RRF Fusion  ──▶  LLaMA 3.3  ──▶  Answer ✓
               Sparse Search (keywords) ──┘

🧠 Why Hybrid RAG?

Dense Only Sparse Only Hybrid RAG
Semantic understanding ✅ ❌ ✅
Exact keyword match ❌ ✅ ✅
Handles typos ✅ ❌ ✅
Technical terms ❌ ✅ ✅
Overall accuracy 🟡 Good 🟡 Good 🟢 Best

🏗️ Architecture

┌─────────────────────────────────────────────────────┐
│                   INGESTION PIPELINE                 │
│                                                      │
│  PDF/TXT ──▶ Chunker ──▶ BGE Embeddings ──▶ ChromaDB│
│                    │                                 │
│                    └──────────────▶ BM25 Index       │
└─────────────────────────────────────────────────────┘
                          │
                          ▼
┌─────────────────────────────────────────────────────┐
│                   QUERY PIPELINE                     │
│                                                      │
│  Query ──▶ Dense Retrieval (top-5)  ──┐              │
│       └──▶ Sparse Retrieval (top-5) ──┴▶ RRF ──▶ LLM│
└─────────────────────────────────────────────────────┘

Stack (zero LangChain):

  • 🔍 ChromaDB — persistent vector store
  • 🧮 BAAI/bge-base-en-v1.5 — sentence embeddings
  • 📊 BM25Okapi — keyword retrieval
  • 🔀 RRF — Reciprocal Rank Fusion (custom implementation)
  • ⚡ Groq API — blazing fast LLM inference
  • 🦙 LLaMA 3.3 70B — answer generation
  • 🖥️ Streamlit — interactive UI

🚀 Quick Start

# 1. Clone & setup
git clone https://github.com/SHALINISAURAV/hybrid-rag
cd hybrid-rag
python -m venv venv && source venv/bin/activate

# 2. Install
pip install -r requirements.txt

# 3. Add your Groq API key
echo "GROQ_API_KEY=your_key_here" > .env

# 4. Drop PDFs in docs/ folder, then index
python ingest.py

# 5. Launch
streamlit run app.py

🔑 Get your free Groq API key at console.groq.com

📁 Project Structure

hybrid_rag/
│
├── 📂 docs/              ← drop your PDFs here
├── 📂 chroma_db/         ← auto-generated vector store
│
├── 📄 ingest.py          ← document ingestion pipeline
├── 📄 retriever.py       ← hybrid retrieval + RRF fusion
├── 📄 app.py             ← streamlit UI + RAG chain
├── 📄 .env               ← GROQ_API_KEY
└── 📄 requirements.txt

✨ Features

  • 📤 Upload PDFs directly from the UI — no manual file moving
  • 🔀 True Hybrid Search — dense + sparse with RRF fusion
  • ⚡ Groq-powered — sub-second LLM responses
  • 📍 Source citations — see exactly which page answered your question
  • 💬 Chat memory — multi-turn conversation within a session
  • 🚫 No LangChain — lightweight, no dependency hell

🛠️ How RRF Works

Reciprocal Rank Fusion merges two ranked lists into one superior list:

# For each document in either list:
rrf_score = 1/(60 + rank_in_dense) + 1/(60 + rank_in_sparse)

# Higher score = more relevant
# Documents appearing in BOTH lists get a big boost

This means a document that ranks #3 in dense AND #2 in sparse will beat a document that ranks #1 in only one list.

📊 Performance

  • ⚡ Query response time: ~1-2 seconds
  • 📄 Handles documents up to 500+ pages
  • 🎯 Chunk size: 512 tokens with 50-token overlap

Built with ❤️ — no hallucinations guaranteed

Try Live Demo