Embeddings That Match Your Domain, Not Just a Model Card

Vector search fails in boring ways: chunks too large, titles lost, tables flattened into soup, acronyms colliding with English synonyms.

Start with content structure. Code, policies, and notebooks want different splitters than blog posts.

Store metadata you will filter on later — product, language, owner, last updated. Filtering before fusion often beats pure cosine ranking.

Test with domain queries, not Wikipedia-style trivia. If your corpus is internal tickets, evaluate on tickets.

Hybrid search is not a fashion statement. Sparse signals still rescue exact identifiers dense models smear.