Date: 2026-01-30
Time: 15:10–16:00
Room: Ballroom A
Level: Intermediate
In this talk I will explore the popular the pgvector HNSW index and show how it plugs into PostgreSQL. We will look at some of the ideas from academic papers about the wider "approximate nearest neighbours" family of indexes to get some basic intuitions about the problem domain, and the basic implementation tricks in comparable systems. Then we'll hunt down as many low-level inefficiencies as we can find along the way as we try them out. The adventure will involve prototyping and evaluating changes to both PostgreSQL and pgvector, including:
Note: despite choosing pgvector as a subject, this is definitely not a talk about AI! It's a PostgreSQL hacking talk. Many of the ideas discussed apply to other kinds of indexes, but this is an interesting motivating example.