Embeddings capture meaning, which is exactly the problem when the words themselves are the point: names, codes, numbers, short phrases. Asked for "one hour", the embedding model returns a shipping chunk about tracking and timing — vaguely time-related — instead of the password page that literally says "a link that works for one hour". That's the real model's behaviour, not a simplification.
Hybrid search runs both kinds and merges the rankings, here with reciprocal rank fusion: each list votes for its results, weighted by rank. Keyword search catches the exact phrase; vector search catches the paraphrases. Most production systems do both.
Try it
Compare all three modes.