Why not both? I wonder if an LSH-based IO-optimised algorithm would work well.

→ View original post on X — @jeremyphoward