Comment by snowstormsun

Isn't one big problem of vector search that it is bounded by the complexity of the knn problem in multiple dimensions, whereas keyword search like bm25 is very fast and scalable because of inverted indices? Annoy and others need to sacrifice accuracy for speed I believe. Hybrid search can be used to first filter a smaller number of candidates and then rerank them using vector search. Performance wise, it is naturally pretty slow and hard to scale compared to keyword search.