Notes from the filter-native frontier.
Engineering posts, retrieval research, and customer dispatches from the team building vector search around the questions you actually ask.
Everything we've shipped and learned.
- 006
AVX-512 Is a Gift to Vector Search, If You Use It Wisely
Two small kernels, a lot of lanes, and one reminder that wider does not automatically mean faster.
11 min read→ - 005
What happens to vector search when 99.99% of the data is filtered out?
Why we built SABLE, a filter-native retrieval engine designed around selectivity from the beginning.
12 min read→ - 004
Introducing SABLE: A Filter-Native Engine for Billion-Scale Vector Search
Most vector databases bolt filtering on after the fact. Production SABLE uses filter-native planning, graph-free ADC over compressed vectors, and exact reranking to keep filtered search fast.
14 min read→ - 003
Navigating Spreading-Out Graphs: An Intuitive Guide to NSG
Graph-based vector search sounds intimidating, but the core idea is elegant. NSG builds a sparse, well-spread graph that lets you navigate to any nearest neighbor in a handful of hops.
10 min read→ - 002
Why "Adding Disk" Isn't an Architecture
Every memory-first vector database is announcing disk support. But bolt-on tiering solves a cost problem. It does not resolve the latency, write-throughput, and filtered-query problems that disk-resident workloads actually surface in production.
10 min read→ - 001
IVF-OPQ Explained: Route, Rotate, Compress, Search
IVF-PQ is the workhorse of large-scale vector search. OPQ makes it dramatically better by learning a rotation that aligns your data with the quantizer. Here's how both work and why OPQ's trick is so clever.
11 min read→