Jonathan S. Katz

> Is there any plan to support these kind of use cases, where combining traditional and hnsw indices would lead to much better performance? I have a POC that does...

One thought around this specific to HNSW: perhaps there's a way towards a "multi-column" style index that would require more links, but allow for this type of query. Let's say...

I'm not suggesting a B-tree (though perhaps we need some semantics like that at the top layer?), but rather, when we compute the distances between vectors, we compute two sets:...

Forgot to post this the other day, but this is what I was thinking around HNSW -- effectively keeping two pairs of links in a "multi-column" style index. The first...

@fanfuxiaoran Thanks for sharing. I haven't read through the doc yet, but on the surface, I'm not sure how that's different from the bitmap index scan and the same issues...

I thought about an approach for `ivfflat` using a multicolumn index strategy. This will example will use two columns: (`embedding`, `category_id`): ## Build We would keep two structure: * The...

@knizhnik It may expand the search radius, but we may not be seeing the best available vectors for a filter. For example, let's say that vectors with a specific filter...

> I think storing the values in the lists themselves in a B-tree structure would have a bad performance as the size of the vector struct is big and we...

@vincenzon How many rows are in the table, and how many `lists` did you use to create the index?

I want to work on benchmarking over the coming days. I do think there's merit to being explicit about SIMD, but I'm not convinced on adding the complexity based on...