📓 NotebookLM Clone — RAG Application

A full RAG pipeline application where users can upload documents and chat with them.

🔗 Live Demo: rag-notebook-lm.netlify.app

---

Advanced RAG Pipeline

Step	Stage	Description
01	Ingestion	User uploads a PDF or .txt file. Text is extracted from PDFs page-by-page using pdfjs-dist to optimize memory usage.
02	Chunking	Custom Recursive Character Splitter — 800-character chunks with 150-character overlap. Splits recursively on separators (paragraphs, newlines, sentences, and spaces) to preserve context. Each chunk includes metadata: { source, chunkIndex, totalChunks, charCount }.
03	Embedding	HuggingFaceInferenceEmbeddings via @langchain/community (using model sentence-transformers/all-MiniLM-L6-v2) to embed chunks in parallel batches.
04	Storage	Pinecone serverless index. Chunks are stored in a dedicated document-level namespace to ensure clean isolation between different uploads.
05	Advanced Retrieval	1. Query Expansion: The original question is expanded into 3 variants (1 original + 2 rephrasings) using gpt-oss-120b via OpenRouter to maximize recall. 2. Multi-Query Search: Similarity search (Top-K=5) is performed for all 3 variants in Pinecone, and candidate results are merged and deduplicated. 3. LLM Reranking (Cross-Encoder): Merged candidate chunks are evaluated in parallel by gpt-oss-120b to score their relevance (0–10). High-scoring chunks are sorted (falling back to cosine similarity on ties) and the top 5 are selected.
06	Grounded Generation	The top 5 reranked contexts are formatted into a grounded prompt. An LLM (gpt-oss-120b:free via OpenRouter) generates the final answer strictly adhering to the context at a low temperature of 0.1 to eliminate hallucinations. Multi-turn chat history is also incorporated.

---

Tech Stack

Backend

Package	Role
express	REST API server
multer	File upload handling
pdfjs-dist	PDF text extraction (page-by-page)
@langchain/community	Hugging Face Embeddings integration
@pinecone-database/pinecone	Vector database for namespace storage
openai	Client for OpenRouter (gpt-oss-120b:free) for generation, query expansion, and reranking

Frontend

Package	Role
vite + react	Fast dev build
tailwindcss	Utility-first styling
UploadZone	Drag & drop file upload UI
ChatWindow	Conversation display
MessageBubble	Message rendering
SourceChips	Citation/source display

---

Setup

Environment Variables

backend/.env

PINECONE_API_KEY=your_pinecone_key
PINECONE_INDEX_NAME=notebooklm
OPENROUTER_API_KEY=your_openrouter_api_key
HUGGINGFACEHUB_API_KEY=your_hf_access_token

frontend/.env

VITE_API_URL=backend_deployed_url

Run Locally

# Backend
cd backend
npm install
npm start

# Frontend
cd frontend
npm install
npm run dev

---

File Structure

notebooklm/
├── backend/
│   ├── src/
│   │   ├── index.js          # Express server
│   │   ├── chunker.js        # Recursive character chunker
│   │   ├── embedder.js       # Hugging Face embedding generator
│   │   ├── vectorStore.js    # Pinecone namespace integration
│   │   ├── reranker.js       # LLM cross-encoder reranking logic
│   │   └── rag.js            # Advanced RAG pipeline orchestrator
│   ├── uploads/              # Temp file storage
│   ├── package.json
│   └── .env
└── frontend/
    ├── src/
    │   ├── App.jsx
    │   ├── components/
    │   │   ├── UploadZone.jsx
    │   │   ├── ChatWindow.jsx
    │   │   ├── MessageBubble.jsx
    │   │   └── SourceChips.jsx
    │   └── main.jsx
    ├── index.html
    ├── package.json
    └── vite.config.js