Lumi — AI-Powered PDF Chat

Lumi is a full-stack RAG (Retrieval-Augmented Generation) app that lets you upload a PDF and chat with it using Google Gemini. Ask questions about your document and get accurate, source-cited answers in real time.

---

Tech Stack

Layer	Technology
Frontend	React + Vite + Tailwind CSS
Backend	FastAPI (Python)
LLM	Google Gemini 2.5 Flash
Embeddings	Google Gemini Embedding 001
Vector Store	ChromaDB
PDF Parsing	PyMuPDF

---

Features

• Upload any PDF and process it instantly
• Ask questions in natural language
• Streaming responses token by token
• Source citations with page numbers
• Multiple chat modes
• Persistent vector storage

---

Project Structure

pdf-chat-app/
├── backend/
│   ├── models/
│   │   └── schemas.py          # Pydantic request/response models
│   ├── routers/
│   │   ├── chat.py             # Chat endpoint (SSE streaming)
│   │   └── upload.py           # PDF upload endpoint
│   ├── services/
│   │   ├── chat_service.py     # LLM streaming logic
│   │   ├── pdf_service.py      # PDF parsing and storage
│   │   └── rag_service.py      # ChromaDB store and retrieval
│   ├── utils/
│   │   ├── chunker.py          # Text splitting
│   │   ├── embedder.py         # Gemini embedding calls
│   │   └── prompt_builder.py   # System and user prompt construction
│   ├── config.py               # Settings via pydantic-settings
│   └── main.py                 # FastAPI app entry point
├── frontend/
│   ├── src/
│   │   ├── api/client.js       # API calls to backend
│   │   ├── components/         # React UI components
│   │   ├── hooks/useChat.js    # Chat state and streaming hook
│   │   ├── pages/              # HomePage and WorkspacePage
│   │   └── store/chatStore.js  # Global state
│   └── vite.config.js
├── requirements.txt
└── README.md

---

Local Development

Prerequisites

• Python 3.11+
• Node.js 18+
• Google Gemini API key — get one free at aistudio.google.com

1. Clone the repo

git clone https://github.com/OnlineBunker/lumi.git
cd lumi

2. Set up the backend

cd backend
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r ../requirements.txt

Create a .env file inside backend/:

GEMINI_API_KEY=your-gemini-api-key-here
LLM_MODEL=gemini-2.5-flash
EMBEDDING_MODEL=gemini-embedding-001
UPLOAD_DIR=./uploads
VECTOR_STORE_DIR=./vector_store
CHUNK_SIZE=500
CHUNK_OVERLAP=100
TOP_K_CHUNKS=5

Start the backend:

python -m uvicorn main:app --reload

Backend runs at http://localhost:8000 Swagger docs at http://localhost:8000/docs

3. Set up the frontend

Open a new terminal:

cd frontend
npm install
npm run dev

Frontend runs at http://localhost:5173

---

Deployment

Frontend — Vercel

1. Push code to GitHub
2. Go to vercel.com → New Project → Import repo
3. Set Root Directory to frontend
4. Deploy — you get a live URL instantly

Backend — Render

1. Go to render.com → New Web Service → Connect repo
2. Set Root Directory to backend
3. Set Build Command to pip install -r ../requirements.txt
4. Set Start Command to uvicorn main:app --host 0.0.0.0 --port $PORT
5. Add environment variables in the Environment tab:
   • GEMINI_API_KEY
   • LLM_MODEL = gemini-2.5-flash
   • EMBEDDING_MODEL = gemini-embedding-001
6. Deploy

Note: Render's free tier spins down after 15 minutes of inactivity. First request after inactivity takes ~30 seconds to wake up.

---

Environment Variables

Variable	Description	Default
GEMINI_API_KEY	Google Gemini API key	required
LLM_MODEL	Gemini model for chat	gemini-2.5-flash
EMBEDDING_MODEL	Gemini model for embeddings	gemini-embedding-001
UPLOAD_DIR	Directory to store uploaded PDFs	./uploads
VECTOR_STORE_DIR	Directory for ChromaDB	./vector_store
CHUNK_SIZE	Characters per text chunk	500
CHUNK_OVERLAP	Overlap between chunks	100
TOP_K_CHUNKS	Chunks retrieved per query	5

---

How It Works

1. Upload — PDF is parsed page by page using PyMuPDF
2. Chunk — Text is split into overlapping chunks using LangChain text splitter
3. Embed — Each chunk is embedded using Gemini Embedding and stored in ChromaDB
4. Query — User question is embedded and compared against stored chunks via cosine similarity
5. Generate — Top matching chunks are injected into a prompt and streamed through Gemini Flash
6. Stream — Tokens arrive in real time via Server-Sent Events (SSE)

---

License

MIT