In real vector search systems, performance is dominated by combining it efficiently with filters. Few test this properly. 🧵
Sep 5, 2025 · 1:54 PM UTC
1
5
21
6,711
What you need to test is performance AND recall at various filter strengts. The challenging area is around 80-99% filter strength, and the devil has made it so that this is where most real-world applications live.
1
1
3
245
We have introduced new tuning parameters in Vespa that lets you improve recall and cost, inspired by Acorn and beam search papers. They *really* help.
1
1
4
257
To optimize all parts of the filter space (make all queries efficient), you need to combine different strategies
1
2
3
363
We'll soon update defaults to give everybody improved performance with no effort, but to really get the best performance you should tune to your case. We have made a guide for that: blog.vespa.ai/tweaking-ann-p…
1
1
7
399
You can also read a full explanation of what these parameters do here: blog.vespa.ai/additions-to-h…
2
7
370