Contributions

I designed and built the Unofficial Guide as an end-to-end retrieval-augmented generation system in Python. The system answers questions about campus life from a corpus of student guides, housing documents, and FAQs, and is built to decline rather than guess when the corpus does not cover a question.

Because the corpus mixed prose, structured FAQ entries, and tabular data, I chose recursive content-aware chunking over fixed-size splitting, producing 263 chunks. Retrieval uses sentence-transformer embeddings with a persistent ChromaDB vector store configured for cosine distance.

Generation is grounded strictly against retrieved context, with a distance threshold that triggers a refusal when nothing relevant is found. I built source attribution programmatically from chunk metadata rather than having the model cite its own sources, so citations cannot be fabricated.

I evaluated the system on a labeled question set and traced two retrieval failures to their causes, fixing one by embedding metadata alongside chunk text and the other by splitting FAQ content per question. A third failure requiring aggregation across the corpus I documented rather than fixed, since it needs metadata pre-filtering ahead of semantic ranking.

Implementation was done with an AI coding assistant working from my specification. The chunking strategy, retrieval configuration, attribution design, and both fixes were my decisions.