Personal project
DocChat
AI Document Intelligence
DocChat is a full-stack document intelligence app. You upload PDFs, DOCX, text files or images to a notebook, ask questions, and get answers with cited source passages through retrieval-augmented generation (RAG).
What he built
- A RAG pipeline on LangChain and pgvector (HNSW index, cosine distance) with Gemini embeddings, and Groq Llama 3.3 with a Gemini fallback through the Vercel AI SDK, streaming each answer token by token.
- Studio tools generated from the documents: flashcards, quiz, report, mind map, data table, slide deck and audio overview.
- Notebook sharing with view-only or chat links, group roles, and exports to Markdown or JSON.
- Supabase Auth (email, magic link, GitHub and Google OAuth), row-level security, rate limits, prompt-injection defenses and CSP headers.
- English and Hebrew with right-to-left support, over 680 Vitest tests and Playwright end-to-end tests in CI.
Stack: Next.js 16, React 19, TypeScript, LangChain, pgvector, Supabase, Vercel AI SDK, Docker
- Next.js
- TypeScript
- RAG
- Supabase
- pgvector
- Vercel AI SDK
