Finally here: MariaDB Vector Preview!

Written by Written by Kaj Arnö Leave a comment on Finally here: MariaDB Vector Preview!
The MariaDB Sea Lion with Victor Vector

We’re here, we’re open source, and we have RDBMS based Vector Search for you! With the release of MariaDB 11.6 Vector Preview, the MariaDB Server ecosystem can finally check out how the long-awaited Vector Search functionality of MariaDB Server works. The effort is a result of collaborative work by employees of MariaDB plc, MariaDB Foundation and contributors, particularly from Amazon AWS. 

Previously on “MariaDB Vector”

If you’re new to Vector, this is what’s happened so far:

The main point: MariaDB Vector is ready for experimentation 

The main points are these

  • AI and LLMs are key, and using them belongs in a “real”, classic, familiar relational database – where the rest of your data is
  • Our vector search implementation is based on the industry standard HNSW algorithm
  • You can start experimenting with AI apps, vectors generated by your favourite AI model (OpenAI GPT-4, Claude, LLama2, etc.) all can be stored and searched for in MariaDB.
  • We have syntax for easily storing vectors (dedicated functions VEC_ToText and VEC_FromText)
  • We will soon blog about benchmarks, but early indications point towards fast retrieval
  • All of MariaDB Vector is on the way of becoming part of standard MariaDB Server
  • You can start experimenting with AI apps 
  • You can download it in source form at https://mariadb.org/download/?p=mariadb&r=11.6.0+Vector  
  • We have prepared a special container image, with CPU optimisations enabled at quay.io/mariadb-foundation/mariadb-devel:11.6-vector-preview
  • We are happy to take contributions, and ready to do so

The fine print: It’s still Alpha

The fine print is this

  • We are talking a Preview, so we don’t provide full binary packages
  • We are talking Alpha release software, which means it may crash
  • We are still extending the syntax
  • We plan a feature complete MariaDB 11.6 Vector Edition release within a few months
  • INSERT speed still needs optimisation
  • The Vector documentation in the Knowledge Base is still work in progress 

Releasing new functionality is always a balance between giving early access to get early feedback, and maturing the software to be stable enough so that testing is meaningful. Given the plentiful requests for testing after the first glimpses we gave at FOSDEM in February, we could no longer wait to get out in the open.

How do we compare our Vector support to that of MySQL Server?

What makes us different from the Vector support in MySQL Server – called Heatwave – is that we have vector search as Open Source. In fact, unlike MySQL, we still lack a separate Vector data type, but we will add it before the full release.

How do we compare our Vector support to that of Postgresql?

What makes us different from the Vector support in PostgreSQL – PG Vector – is that we are looking at helping not just the MariaDB Server users but also the MySQL Server users to get easy Open Source access to Vector functionality. We’ve always made the point that migrating from MySQL Server to MariaDB Server is dead easy, and getting access to Vector functionality is a great reason to get going.

We’re also different from PG Vector in that MariaDB Vector isn’t a plug-in, but a part of core MariaDB Server. We plan MariaDB 11.7 to be the first “normal” release to contain the Vector functionality. 

What makes us similar to PG Vector is our interest and capability of working with contributors. We are happy to note that already the first version of MariaDB Vector has significant contributions, where I want to particularly highlight code by Hugo Wen of Amazon AWS. Working with AWS has helped us tune our contribution and review processes. 

How to get started