Paragraphs are chunked at build time, embedded with a sentence-transformer, and projected to 2D with UMAP. Search loads the same model into your browser via transformers.js and ranks chunks by cosine similarity — a fully client-side retrieval pipeline.