Date: 2025-05-15
Time: 13:00–13:50
Room: Breakout Room 1
Level: Advanced
Generative AI and associated technologies made modern vector search an integral part of building applications, such that many traditional databases have implemented a vector search feature. While PostgreSQL itself doesn't have a vector search capability for the size of vectors from modern ML models, PostgreSQL's extension framework has let developers build out these features. pgvector, the most popular of these extensions, has helped establish PostgreSQL as a "vector database" while itself becoming a benchmark on both capabilities and performance of vector search.
Over the past four years, pgvector has built out the core capabilities required to support vector search, with benchmark showing that it is amongst the leading solutions available for vector search. However, the characteristics of searching for vectors has pushed some PostgreSQL capabilities to their limits, and the next set of innovations in vector search for PostgreSQL need to happen in PostgreSQL itself, not extensions.
In this talk, we'll review the pgvector implementation and how it has evolved to date. We'll cover some recent PostgreSQL features and how they can be applied to improve vector search in pgvector. Finally, we'll review specific parts of PostgreSQL (e.g. TOAST, page size, I/O), how they currently impact vector search, and how we can improve PostgreSQL to further advance this domain.