Unlike text search, filtered vector search can be slower

Full-text search instincts about filters don’t translate to vector search

In full-text search, the rule has been to filter before ranking. Make results more precise while speeding up search.

While filters do improve precision, they don’t speed up graph based vector search: in fact search gets slower.

Navigating vector spaces is like navigating a map. Except in more dimensions. You want to find nearest neighbors to an address? OK, it’s not hard to build a data structure that says “Here’s Doug’s house and 10 nearby neighbors”.

But typical search is different. You frequently say “I want 10 of Doug’s neighbors, but only blue houses”. Well, you break the data structure. You need to explore neighbors, and neighbors-of-neighbors, gathering houses until you get 10 nearby blue neighbors.

So be careful — adding a filter can vector search slower! More conditions, deeper graph traversal.

-Doug

​

​

This is part of Doug’s Daily Search tips - subscribe here