Campus Handbook Assistant
A retrieval-augmented chat tool that answers questions about registration, graduation requirements, and department rules from official PDFs — with citations.
Implementation
Built after watching the same handbook questions get re-asked in every department group chat every semester. PDFs from the registrar's office are chunked, embedded, and indexed; a FastAPI backend retrieves relevant chunks and passes them to an LLM to generate an answer that always includes a citation back to the source document and page. If retrieval confidence is low, it says so rather than guessing.
Architecture
Document ingestion pipeline: PDF → text extraction → semantic chunking (~500 tokens, overlap 50) → embeddings stored in pgvector. Query path: user question → embedding → top-k retrieval → reranking → prompt assembly with citations → LLM generation. The frontend is a plain Next.js chat UI; there's no conversation memory across sessions by design, to keep answers grounded in the current question rather than drifting with chat history.
Technologies
Known limitations
- Source PDFs are re-ingested manually when the registrar publishes updates — there's no automatic change detection yet.
- Retrieval quality drops noticeably on tables and multi-column PDF layouts, which the chunker doesn't parse well.
- Korean and English handbook versions are indexed separately and can occasionally give slightly different answers to the same question.
Members
- Yuna Kim
- Minji Lee
- Ethan Yoon