The problem
Long PDFs are difficult to search when you're studying. My friend was repeatedly going through notes and study material to find specific information.
I wanted to build something more useful than a generic chatbot:
What if the AI could answer questions from the study material — but refuse to answer when the material doesn't contain enough evidence?
What I built
LearnLens lets a user upload study PDFs and ask questions about them.
It:
- extracts text while preserving page boundaries
- chunks the document
- generates BGE embeddings
- stores vectors in PostgreSQL + pgvector
- retrieves relevant passages
- gives those passages to Gemma 3 4B
- returns a grounded answer with page-level sources
- refuses to guess when retrieval doesn't provide sufficient evidence The interesting part: "I won't guess" Most AI interfaces make it easy to assume that a confident-sounding answer is correct. LearnLens takes the opposite approach. If the retrieved material isn't sufficient, it responds: I couldn't find enough information in your uploaded study material to answer this reliably. I won't guess.
The system doesn't manufacture a confidence percentage. It uses retrieval evidence to determine whether there is enough supporting material to answer.
Demo
Live frontend: https://learnlens-blond.vercel.app
Demo video: https://youtu.be/TmwEB-ZvwCU
GitHub: https://github.com/shivamsinghx/LearnLens
Prize category- Best Use of Gemma
What's next
- public backend deployment
- more document formats
- better retrieval evaluation
- hybrid retrieval
- additional open-weight models P.S- I'm currently working on it(project in progress)
Top comments (0)