3 ms·
Good to see another engine leverage ScANN outside of Google. HNSW uses lots of RAM and it’s interesting how all the major engines settled on that algo. I’m in
by binarymax 3y ago
Good to see another engine leverage ScANN outside of Google.
HNSW uses lots of RAM and it’s interesting how all the major engines settled on that algo.
I’m interested in how they apply filtering, since with codebook based similarity such as PQ and ScANN it’s not trivial.
Maybe one day we’ll also see someone implement a production ready Vamana engine too, which also does really well at the billion scale.
- fzliu 3y agoAgreed. There's a lot of great index types out there - HNSW is incredible, but algorithms such as ScANN (and PQ) have their place in the ecosystem. Tree-based vector indexes aren't bad either, especially if we can find a way to make the random projections more efficient.