MariaDB Vector Search – The fastest vector search in relational databases

Abstract

Vectors are key to how AI models represent data semantics. Searching for vectors has now become a key requirement of databases in order to facilitate AI first applications. MariaDB Server will soon have Vector support in an LTS version. In this talk we will discuss what MariaDB vector is, how it works behind the scenes, as well as possible use cases and future roadmap.

MariaDB Vector introduces a new high-level interface for indexing within MariaDB Server. This interface allows one to create custom indexing strategies. Vector search requires a special kind of index.

The algorithm that is used by MariaDB and many other vector databases is called Hierarchical Navigable Small Worlds (HNSW). In this talk we will focus on the basis of HNSW, why the algorithm returns approximate results and what impacts its performance (tuning parameters).

We will also discuss the difference between a Generative AI model and an Embedding AI model and how they can be used to build Retrieval Augmented Generation applications using MariaDB as a datastore.

Finally we’ll describe the current eco-system of Vector databases what are their strengths and weaknesses. With the information provided, one will be able to make a more informed decision between choosing a dedicated Vector database, or stick with a traditional Relational Database with vector search support.

Panel

Let’s discuss use cases with MariaDB Vector.
What kind of applications do you want to build that MariaDB Vector can enable?

Is it a QA system? Is it a faster or better text search? RAG or just information retrieval?

Should MariaDB Server abstract away some of the complexities of an ML pipeline or is that better delegated to a middleware platform like MindsDB, LLamaIndex, etc.

How important is hybrid search for you? Prefiltering of data, etc.

Vicențiu Ciorbaru, MariaDB Foundation

Vicențiu Ciorbaru is based in Romania, Bucharest. Having had to fix problems in many places, ranging from query-optimiser, authentication, replication, packaging, as well as platform specific problems, Vicențiu has had experience with many parts of the codebase. He is also a fast learner and can jump on any problem that surfaces. Past projects include Roles, Window Functions, Custom Aggregate Functions.