A local-first, full-stack GenAI study assistant with a premium, polished UI. Upload your notes and ask questions — answers are grounded in your actual documents using a RAG pipeline, not hallucinated.
This project is designed to run locally and is not deployed publicly to avoid reliance on paid APIs.
| Login | Chat (Math + Streaming) |
|---|---|
 |
 |
Upload notes → chunked → embedded → stored in FAISS
↓
Ask a question → Smart Router → needs docs? → YES → search FAISS → relevant chunks
→ NO ─────────────────────┐
↓
LLM generates answer (streamed in real-time)
- Upload — PDF / TXT / MD → split into chunks → embedded into vectors
- Ask — Smart router skips RAG for simple questions (math, greetings)
- Retrieve — FAISS finds relevant chunks, irrelevant ones filtered by score
- Generate — LLM streams the answer token-by-token
- Cite — Collapsible source tags show which document was used
| Layer | Technology |
|---|---|
| Frontend | React 19, TypeScript, Vite, Tailwind CSS 4 |
| Rendering | React Markdown, KaTeX (math) |
| Backend | FastAPI, Python 3.12 |
| Auth | JWT + bcrypt |
| Database | SQLite + SQLAlchemy |
| Vector Store | FAISS |
| Embeddings | Ollama nomic-embed-text |
| LLM | Ollama llama3 (default) / Gemini 2.0 Flash (optional) |
| Orchestration | LangChain |
| Streaming | Server-Sent Events (SSE) |
- RAG Pipeline — Answers grounded in uploaded documents
- 100% Local — Runs via Ollama, no API key needed
- Streaming — Tokens appear in real-time
- Smart Routing — Math / greetings skip RAG for instant answers
- Relevance Filtering — Sources shown only when actually relevant
- Auth — JWT login/register with hashed passwords
- Chat History — Multiple saved sessions
- Math Support — LaTeX rendered via KaTeX
- Dual LLM — Optional Gemini with automatic Ollama fallback
StudyGenie/
├── backend/
│ ├── app/
│ │ ├── main.py # FastAPI entry point
│ │ ├── api/endpoints/
│ │ │ ├── auth.py # Login / Register
│ │ │ ├── chat.py # Chat + Streaming + Routing
│ │ │ └── documents.py # Upload / List / Delete
│ │ ├── core/
│ │ │ ├── config.py # Settings (.env)
│ │ │ ├── database.py # SQLAlchemy + DB
│ │ │ └── auth.py # JWT tokens
│ │ └── services/
│ │ ├── ingestion.py # Chunking + processing
│ │ └── rag.py # FAISS + embeddings
│ ├── requirements.txt
│ └── .env.example
├── frontend/
│ ├── src/
│ │ ├── App.tsx
│ │ ├── api/client.ts # API + SSE streaming
│ │ └── components/
│ │ ├── AuthPage.tsx
│ │ ├── ChatInterface.tsx
│ │ └── Sidebar.tsx
│ └── package.json
├── docker-compose.yml
└── README.md
- Python 3.12+
- Node.js 18+
- Ollama
ollama pull llama3
ollama pull nomic-embed-text
ollama servecd backend
python -m venv venv
venv\Scripts\activate # Windows
pip install -r requirements.txt
copy .env.example .env # Windows
uvicorn app.main:app --host 0.0.0.0 --port 8000 --reloadcd frontend
npm install
npm run devOpen the local URL shown in the terminal, register an account, upload your notes, and start asking questions.
For faster cloud-based responses, add a free Gemini API key to backend/.env:
GEMINI_API_KEY=your_key_here
LLM_PROVIDER=gemini
Gemini is used for chat with automatic Ollama fallback if quota is exceeded.
- How RAG pipelines work end-to-end (chunking → embedding → retrieval → generation)
- Local vs cloud LLM trade-offs (latency, quality, cost)
- SSE streaming for real-time token delivery
- Smart query routing to skip expensive operations
- Relevance filtering with vector similarity scores
- Full-stack integration: React + FastAPI + FAISS + LangChain
This project deepened my understanding of latency, model quality trade-offs, real-time streaming systems, and multi-model integration.