What do you expect from vector storage in databases?

We’re no mind readers, so from time to time, we like to do polls. Polls are quantitative in nature, so coming up with the right question is not enough – we need to do a bit of mind reading in coming up with the alternatives.

Quick development of text based RAG apps

Our hypothesis was that RAG is the cool thing to do with vector based databases, and specifically text based RAG. The conference talks we’ve given on MariaDB Vector (such as at the 24th SFSCON in Bozen, Südtirol, Italy on 8 Nov 2024) have stressed the value of easily being able to develop AI applications that answer user prompts based on knowledge in a specific text mass, not on the overall training data of an LLM. This is spot on what you can do with MariaDB Vector.

Our poll more or less confirms the hypothesis, with 30% (of the “huge” base of 20 respondents) picking “Fast text RAG dev” as the top response.

Fast response times

MariaDB Server has always taken pride in the trinity of stability, performance, and ease of use. Meaning: We take performance seriously. Nearly half of the respondents – 45 % – put “Fast response times” on spot one. What a relief, then, that even the initial release of MariaDB Vector fared well in the area of performance – as witnessed by the blog entry How Fast Is MariaDB Vector? by Sergei Golubchik.

Integration in a standard (non-vector) RDBMS

MariaDB Vector is a piece of functionality in a standard Open Source Relational DBMS. It’s not a one-trick pony good only for Vectors. Its claim is to deliver the vector functionality needed, with a fast response time, while still retaining all the good stuff offered by MariaDB Server. A standard database has the syntax we all know and love, plenty of functionality out of the box, lots of tools, a pool of experts, and the functionality to combine source data, vector data, any relational data in the same query.

Of the poll respondents, 20% rated “Leverage standard RDBMS” their top expectation on a vector database.

Don’t forget images and videos!

To be honest, I haven’t yet seen a MariaDB Vector based app using images, audio, or video. But nothing prevents MariaDB Vector users to create such applications. Binary data can be vectorised and used for nearest-neighbour search, just as well as text. “Fast multi-modal RAG dev” got a 5% response in our poll, so there must be others thinking along the same lines.

Conclusion: “I told you so!”

I got what I expected from our poll: A few answers confirming our hypothesis that MariaDB Foundation’s LinkedIn followers primarily want to be able to quickly develop text based RAG applications, with fast response times. Getting rough validation of our thinking is still hugely important, to know that we’re on track in making MariaDB Server the default database to store your vectors. Our stake is to become a standard component in AI apps.