📓 NotebookLLM

Chat with your documents — with answers that verify themselves.

A NotebookLM-style RAG app where you upload a PDF, TXT, or CSV and have a real conversation with it. Every answer is grounded strictly in your document and graded by an LLM-as-judge before you ever see it.

Live Demo →

---

✨ Why it's different

This isn't a basic "embed → search → answer" RAG. It's an advanced pipeline that cleans up the question, searches smarter, and checks its own work:

Feature	What it does
🌍 Multilingual	Ask in any language — the question is translated to English for retrieval and the answer is translated back.
✍️ Query rewriting	Turns a follow-up like "and what about its cost?" into a standalone search query using chat history.
🧩 Sub-query decomposition	Splits a complex question into focused sub-questions, retrieves for each, then merges and de-duplicates.
⚖️ LLM-as-judge	Scores every answer for groundedness and relevance; weak answers are auto-regenerated, and the verdict is shown in the UI.
🔒 Strictly grounded	The model answers only from retrieved context and cites source numbers — no outside knowledge.
🔍 Full transparency	Expand any answer to inspect the rewritten query, sub-queries, judge verdict, and the exact chunks used.

---

🧠 How a question flows

flowchart LR
    Q[User question] --> L{English?}
    L -- no --> T[Detect + translate]
    L -- yes --> R[Rewrite as standalone query]
    T --> R
    R --> D[Decompose into sub-queries]
    D --> S[Parallel vector search in Qdrant]
    S --> M[Merge + de-duplicate top chunks]
    M --> G[Generate grounded answer]
    M --> J1[Relevance check]
    G --> J2[LLM-as-judge: grounded?]
    J2 -- weak --> G2[Regenerate stricter]
    J2 -- good --> A[Answer + verdict + sources]
    G2 --> A

Loading

On upload: the file is loaded, chunked, embedded with HuggingFace, and stored as vectors in Qdrant.

---

🛠️ Tech Stack

Layer	Technology
Frontend	Vite + React + Tailwind CSS
Backend	Node.js + Express
Answer LLM	Groq llama-3.3-70b-versatile
Judge LLM	Google gemini-2.0-flash (falls back to Groq if no Gemini key)
Embeddings	HuggingFace BAAI/bge-small-en-v1.5 (384-dim)
Vector DB	Qdrant Cloud

---

🚀 Quick Start

Prerequisites: Node.js 18+, and free API keys from Groq, HuggingFace, and a Qdrant Cloud cluster. A Google AI Studio key is optional (enables the Gemini judge; otherwise it falls back to Groq).

1. Backend

cd backend
npm install --legacy-peer-deps
cp .env.example .env      # fill in your keys
npm run dev               # → http://localhost:3000

backend/.env:

QDRANT_URL=https://<your-cluster>.cloud.qdrant.io:6333   # ⚠️ include the :6333 port
QDRANT_API_KEY=<your-qdrant-key>
QDRANT_COLLECTION_NAME=<collection-name>
HF_API_KEY=<your-huggingface-key>
GROQ_API_KEY=<your-groq-key>
GEMINI_API_KEY=<your-gemini-key>     # optional
PORT=3000

2. Frontend

cd frontend
npm install
npm run dev               # → http://localhost:5173

frontend/.env:

VITE_API_URL=http://localhost:3000   # point at local backend; omit to use the deployed API

Open the app, upload a document, and start asking. 🎉

---

📡 API Reference

Method	Endpoint	Description
POST	/api/upload	Upload & ingest a document (multipart/form-data, field document)
DELETE	/api/upload	Remove the indexed document
POST	/api/chat	Ask a question — body { query, history }
GET	/api/status	Index readiness and vector count
GET	/api/health	Health check

---

🧱 Chunking Strategy

PDF / TXT — RecursiveCharacterTextSplitter · size 1000, overlap 200 Splits on natural boundaries in priority order (paragraph → line → sentence → word → character), so chunks never break mid-word and the overlap preserves context across boundaries.

CSV — Row-batch chunking · 10 rows per chunk Each row becomes Column: value | Column: value, with column names repeated in every chunk so the model always knows what each value means.

---

📂 Project Structure

NotebookLLM/
├── backend/
│   └── src/
│       ├── config/      → gemini.js · groq.js · qdrant.js
│       ├── routes/      → upload.js · chat.js · status.js
│       ├── services/    → ingestionService · chatService · translator
│       │                  queryRewriting · subQueries · llmJudge
│       ├── middleware/  → errorHandler.js
│       └── index.js
└── frontend/
    └── src/
        ├── components/  → UploadZone.jsx · ChatMessage.jsx
        ├── lib/         → api.js
        └── App.jsx

---

Built with ❤️ as a NotebookLM-style clone · Groq · Qdrant · LLM-as-judge