Free Lesson
Tuning HNSW parameters for filtered search
60 min
Apr 16, 2026 1:00 PM
Virtual (Zoom)
In this video
What you'll learn
Why filters make HNSW vector search slow
We'll start with a short intro on how HNSW works and why filters are expensive and hurt recall.
Parameters to improve performance/recall
Pros and cons of post filters, ACORN-1, brute force kNN, overfetching, and adaptive beam search.
Tools to tune vector search filter search
How HNSWTuner+VespaNNParameterOptimizer allows you to change these knobs and see the impact of latency and recall
Why this topic matters
Most vector search involves filtering, which has a big impact on both performance and quality compared to unfiltered Approximate Nearest Neighbor search on an HNSW. We'll discuss ways to limit this impact, keeping queries nice and fast.
You'll learn from

Radu Gheorghe
Software Engineer, Vespa.ai

Trey Grainger
Founder @ Searchkernel; Author, AI-Powered Search
Previously at
.png&w=1536&q=75)