HNSW became the defacto default specifically because you don’t need to precalculate the index and it updates as writes come in.

This is a great article, and all the points are there, but the truth is that most teams running million scale vectors don’t want the operational complexity of quantizing offline in some frequency. They’ll gladly outsource the costs to paying for RAM instead of some IVFPQ calculation.

However if there were a solution that “just worked” to handle PQ for shards in near real time for updates that also had sane filtering, that would be really nice.