This started as my NTCC project in college, but what really pulled me in was wanting to solve a real problem with LLMs — they can hallucinate or give outdated answers, which is risky for academic content that needs to be accurate and traceable. So I set out to build a tutoring chatbot that retrieves relevant context from a knowledge base before generating a response, grounding answers instead of just guessing plausibly. My approach evolved a lot along the way: I started with local sentence-transformer embeddings but hit RAM limits deploying on Streamlit Community Cloud's free tier, so I switched to the Gemini embeddings API; I had to pin my Python version after dependency conflicts, and debug a Gemini model deprecation mid-project. I also went from zero coding background to learning Git and GitHub through the browser UI, and eventually evaluating my system's output quality with RAGAS instead of just assuming it worked — which changed how I thought about "done."

RAG-based Tutoring Chatbot

A library in your pocket. A tutor by your side.

A library in your pocket. A tutor by your side.