Session brief

Supabase Vector: The Postgres Vector database: Paul Copplestone

Overview

This talk makes a case for using PostgreSQL with the PG Vector extension as a robust solution for storing and querying vector embeddings. It highlights how this approach integrates seamlessly with existing PostgreSQL infrastructure, offering a powerful and extensible option for AI applications, particularly for developers building with open-source tools and services.

Who should watch

• AI Engineers
• Product Managers
• Builders working with vector databases
• Developers seeking to integrate AI capabilities into existing PostgreSQL workflows
• Those interested in open-source AI solutions

Key takeaways

• Supabase offers a backend-as-a-service built on PostgreSQL, including authentication, edge functions, storage, real-time capabilities, and a vector offering powered by PG Vector.
• PG Vector, an open-source PostgreSQL extension, is presented as a viable and performant solution for storing and indexing vector embeddings, supporting AI application development.
• The integration of PG Vector into Supabase accelerated the growth of new databases on their platform, with 10-15% of new databases now utilizing it.
• A demonstration showcased how PostgreSQL's partitioning and triggers can be leveraged to manage and filter vector data, such as segmenting "good" and "bad" cat images based on similarity scores.
• PostgreSQL's extensibility, row-level security, and the ability to store embeddings alongside operational data in a single round trip are highlighted as key advantages.
• While specialized vector databases exist, PostgreSQL with PG Vector is positioned as a strong contender for the core use case of storing, indexing, and fetching embeddings.
• Future development for Supabase Vector includes sharding with the Citus extension to handle billions of vectors and scale transaction throughput linearly.

Notable quotes

Watch on YouTube →

Unofficial community note. Prefer the recording for nuance.