2025-01-14 · Min-jun Park

Shipping Your First RAG Endpoint in Busan

When teams in Dong-gu ask where to start with retrieval-augmented generation, we point them to a narrow first endpoint: one collection, one embedding model, and a single evaluation sheet.

The capstone path in our Generative AI App Integration course sequences chunking experiments in week four before any UI work. Participants log precision@5 on a fixed question set so improvements are visible, not anecdotal.

Mentors review latency and token budgets together. Most first drafts over-fetch context; trimming top-k from eight to four often stabilizes answers without retraining anything.

We do not promise production scale on day one. The goal is a reviewed integration you can demo to stakeholders—with known failure cases documented in the README.