📄 Multi-Tenant RAG Platform (Document Q&A System)

A full‑stack multi‑tenant Retrieval‑Augmented Generation (RAG) platform that allows teams to upload documents (PDF, DOCX, Images), index them into a vector database, and ask questions strictly based on the uploaded content — with streaming AI responses.

This project is built as a real‑world, production‑style system with authentication, file ingestion, OCR fallback, vector search, and LLM integration.

---

🚀 Features

• 🔐 Authentication & Multi‑Tenancy (team‑based isolation)

• 📤 File Upload Support

  • PDF (text + OCR fallback)
  • DOCX
  • Images (PNG / JPG via OCR)

• 🧠 RAG Pipeline (Retrieve → Augment → Generate)

• 📦 Vector Database with Qdrant

• 🤖 Local LLM via Ollama (LLaMA 3)

• ⚡ Streaming AI Responses (token‑by‑token)

• 🧾 Context‑only Answers (No Hallucination)

• 💬 Modern Chat UI (Markdown + streaming cursor)

---

🛠️ Tech Stack

Frontend

• Next.js 14 (App Router)
• React
• TypeScript
• Tailwind CSS
• React Markdown
• Axios
• React Hot Toast

Backend

• Node.js
• Express.js
• TypeScript
• Multer (file uploads)
• JWT Authentication
• Axios

AI & Data

• Ollama (LLaMA 3 for generation)
• Qdrant (vector database)
• OCR: Tesseract.js
• DOCX Parsing: Mammoth
• PDF Handling: Poppler (pdftoppm)

---

🧠 System Architecture

Frontend (Next.js)
     ↓
Express API (Auth + Upload + Chat)
     ↓
Text Extraction (PDF / DOCX / OCR)
     ↓
Embedding Generation
     ↓
Qdrant Vector Store (per team)
     ↓
Context Retrieval
     ↓
Ollama (LLaMA 3)
     ↓
Streaming Answer → Frontend

---

📂 Project Structure

backend/
 ├─ src/
 │  ├─ controllers/
 │  │  ├─ ingest.controller.ts
 │  │  └─ ask.controller.ts
 │  ├─ routes/
 │  │  ├─ ingest.routes.ts
 │  │  └─ ask.routes.ts
 │  ├─ utils/
 │  │  ├─ rag.ts
 │  │  ├─ qdrant.ts
 │  │  └─ token.ts
 │  ├─ app.ts
 │  └─ server.ts
 ├─ uploads/
 └─ .env

frontend/
 ├─ app/
 │  └─ chat/page.tsx
 └─ lib/api.ts

---

⚙️ Environment Variables

Create a .env file in backend/:

PORT=4000
JWT_SECRET=your_secret_key
QDRANT_URL=http://localhost:6333
QDRANT_COLLECTION=docs
OLLAMA_URL=http://localhost:11434

---

🧪 Prerequisites

Make sure you have:

• Node.js 18+
• Ollama installed
• Qdrant running
• Poppler installed (for PDF → image OCR)

Install Poppler (Windows)

• Download Poppler
• Add bin/ folder to PATH
• Verify:

pdftoppm -h

---

▶️ How to Run the Project

1️⃣ Start Qdrant

run qdrant.exe

(or use the official binary)

---

2️⃣ Start Ollama & Pull Model

ollama pull llama3
ollama serve

---

3️⃣ Backend Setup

cd backend
npm ci
npm run dev

Backend runs on: http://localhost:4000

---

4️⃣ Frontend Setup

cd frontend
npm ci
npm run dev

Frontend runs on: http://localhost:3000

---

📤 Uploading Documents

• Endpoint: POST /upload
• Auth required (JWT)
• Form‑Data key: file

Supported formats:

• .pdf
• .docx
• .png, .jpg, .jpeg

Each document is:

• Parsed / OCR‑ed
• Converted into text
• Embedded
• Stored in Qdrant with team isolation

---

💬 Asking Questions

• Endpoint: POST /chat/ask
• Request body:

{
  "question": "What skills are mentioned in the resume?"
}

🔒 Anti‑Hallucination Prompt

The model is instructed to:

• ❌ Never use outside knowledge
• ❌ Never guess
• ✅ Answer only from retrieved context
• ✅ Say "Not found in the context" if missing

---

⚡ Streaming Responses

• Backend streams tokens using res.write()
• Frontend reads via ReadableStream
• Animated inline loader appears during streaming

---

🧠 Example Output

🚨 Answer within the context:

📌 Skills
• JavaScript, TypeScript
• React, Next.js
• Node.js, Express
• Tailwind CSS

---

🛡️ Security & Isolation

• JWT‑based authentication
• Each document stored with team metadata
• Qdrant filters ensure no cross‑team data leaks

---

📈 Future Improvements

• ✅ Real embedding model (nomic‑embed‑text)
• 🔍 Chunking & overlap
• 🗂️ Document management UI
• 📊 Token usage stats
• 🌍 Cloud deployment

---

👨‍💻 Author

Aryan Gawade

• GitHub: https://github.com/NoB0T21
• LinkedIn: https://www.linkedin.com/in/aryan-gawade-3723672ab/

---

⭐ Final Note

This project demonstrates real‑world RAG architecture, AI streaming UX, and production‑ready backend patterns — suitable for internships, final‑year projects, and portfolios.

If you like it, ⭐ the repo!