📄 DocuMind AI

ChatGPT-like PDF & Knowledge Assistant (RAG-Powered)

DocuMind AI is a full-stack Retrieval-Augmented Generation (RAG) application that allows users to upload PDFs and chat with them intelligently, while also supporting general AI conversation when no document context is provided.

Built with a production-grade architecture, DocuMind AI combines modern LLMs, vector search, and a clean chat UI to deliver accurate, context-aware answers.

🚀 Live Demo

• Frontend (Vercel): https://docu-mind-ai-zeta.vercel.app/

• Backend API (Render): https://documind-ai-2.onrender.com

✨ Key Features

📌 Document-Aware Chat (RAG)

• Upload any PDF document

• Automatically chunks and indexes content

• Answers questions strictly from document context

🤖 General AI Chat

• Works like ChatGPT when no PDF is selected

• Intelligent fallback if document context is not relevant

🧠 Contextual Memory

• Maintains conversation history per session

• Supports natural follow-up questions

⚡ Scalable Vector Search

• Uses Pinecone for fast semantic similarity search

• Isolates documents using namespaces (multi-PDF ready)

🌐 Modern Web UI

• ChatGPT-style interface

• PDF upload + chat in one flow

• Clear distinction between PDF mode and General mode

☁️ Cloud Deployed

• Frontend on Vercel

• Backend on Render

• Designed for real-world production usage

🏗️ System Architecture

User (Browser)
   ↓
React UI (Vercel)
   ↓
Express API (Bun, Render)
   ↓
LangChain RAG Pipeline
   ↓
Pinecone Vector Database
   ↓
LLM (Groq)

🧠 How It Works (High Level)

1. PDF Upload

   • User uploads a PDF

   • Backend saves it temporarily

   • Document is chunked and embedded

   • Vectors are stored in Pinecone under a unique namespace

2. Chat Flow

   • User asks a question

   • If a pdfId exists → RAG search is performed

   • Relevant chunks are sent to the LLM

   • If no relevant context is found → fallback to general chat

3. Conversation Memory

   • Recent messages are stored per session

   • Enables follow-up questions with context

🛠️ Tech Stack

Frontend

• React (Vite)

• JavaScript

• Fetch API

• Deployed on Vercel

Backend

• Bun runtime

• Node.js / Express

• LangChain

• Multer (PDF upload)

• Deployed on Render

AI & Data

• Groq LLM API

• Pinecone Vector Database

• Google Generative AI Embeddings

• LangChain PDF Loader

📂 Project Structure

company-chatbot/
├── routes/
│   ├── chat.route.js
│   └── upload.route.js
├── services/
│   ├── chat.service.js
│   └── pdf.service.js
├── prepare.js
├── server.js
├── uploads/
└── package.json

Frontend is maintained separately in a chat-ui project.

⚙️ Environment Variables

Backend (Render)

GROQ_API_KEY=your_api_key
PINECONE_API_KEY=your_api_key
PINECONE_INDEX_NAME=your_index_name

Frontend (Vercel)

VITE_BACKEND_URL=https://your-backend-url

▶️ Running Locally

Backend

bun install
bun server.js

Frontend

npm install
npm run dev

🔐 Design Decisions & Stability Notes

• Uses disk-based PDF uploads (safe on Render)

• Avoids unstable PDF parsing logic

• Relies on LangChain’s proven PDFLoader

• Keeps infrastructure simple and reliable

• Optimized for correctness before optimization

📈 Future Possible Enhancements

• User authentication (JWT / OAuth)

• Persistent chat history (Redis / DB)

• Streaming responses (typing effect)

• PDF management dashboard

• Usage analytics & limits

🧑‍💻 Author

Savan Kansagara GenAI & Full-Stack Developer

• GitHub: https://github.com/savankansagara1

• LinkedIn: https://www.linkedin.com/in/savan-kansagara/