IntelliDocs

Reading through long PDFs to answer specific questions doesn't scale, and asking an LLM directly — without grounding — risks confident-sounding but incorrect answers with no way to verify the source.

Built a multi-user RAG platform where every answer is grounded in retrieved, embedded chunks of the user's own documents, with page-level citations so the answer can be checked against the source PDF directly, not just trusted at face value.

1. 01Flask backend (Flask-JWT-Extended, Flask-CORS) handling auth, document management, and chat endpoints
2. 02PyMuPDF extracts text from uploaded PDFs, followed by a smart-chunking step
3. 03Jina AI generates embeddings for each chunk, indexed in FAISS for similarity search
4. 04Retrieved chunks are assembled into a grounded prompt and passed to Groq's LLM for the final answer
5. 05PostgreSQL (Neon) via SQLAlchemy stores users, documents, and conversation history
6. 06Dockerized and deployed on Render

• ▸Implemented JWT-based multi-user auth with per-user document isolation at the storage layer
• ▸Built the extraction → chunk → embed → index pipeline so new PDF uploads become searchable immediately
• ▸Tied citation metadata to chunk provenance so every answer can point back to an exact page
• ▸Containerized the app with Docker and deployed it to Render against a managed Neon Postgres instance

• ▸Keeping embeddings CPU-optimized to control cost while still deploying on a lightweight production image
• ▸Designing chunk-to-citation tracking so a generated answer's sources remain accurate after retrieval and generation

A live, publicly deployed multi-user RAG platform with page-level citations, running in production on Render with a managed Postgres database.