Episode

Vector data in Postgres - how's it different from 'normal' data?

with Heikki Linnakangas

POSETTE: An Event for Postgres 2024

Video of a conference talk about vector data in Postgres presented by Heikki Linnakangas at POSETTE: An Event for Postgres 2024. AI works with vectors. But what is a vector? How is a vector in Postgres different from integers, text, geospatial data, or other traditional Postgres datatypes? Why can't you create a regular B-tree index over vectors? Ok, but why not a GiST index?

Heikki will touch on these Postgres and pgvector topics:

  • pgvector extension's data types, including the new sparse and bit vector types
  • Curse of dimensionality with vector databases
  • Vector search algorithms, including HNSW (hierarchical navigable small worlds)

Heikki is a cofounder of Neon and a long-time PostgreSQL developer and committer. Since 2023, he has also been contributing to pgvector.

Chapters

  • 00:00 - Intro
  • 01:37 - Vector search is expensive
  • 02:30 - What is a vector?
  • 03:57 - Vectors are large
  • 06:36 - What is vector search?
  • 11:45 - Vector search is approximate
  • 15:17 - Types of vector search algorithms
  • 20:12 - Three new datatypes in pgvector 0.7.0
  • 23:55 - Summary

Connect

Azure
Azure Database for PostgreSQL