Andre Zayarni: Powering Generative AI with Qdrant’s Scalable Vector Database

Machine learning models do not process text in the form of letters like humans do. Instead, they represent text just like any other data as vectors – columns of numbers that represent features – since vectors are easier for them to handle.

While you could store vectors in a normal database, dealing with vectors has become such an integral part of machine learning that startups like Qdrant have built databases specifically to store and manage high-dimensional vectors efficiently.

Two years ago, we interviewed Andre Zayarni, the co-founder and CEO of Qdrant, for the first interview on this blog about their founding story and technology. Since then, Qdrant has built a powerful vector database and similarity search engine, tailored for machine learning. While the database is open-source and free to use, Qdrant has recently launched its fully managed cloud service, getting developers up and running in a breeze.

Following its latest $28M Series A funding round, led by Spark Capital and joined by existing investors Unusual Ventures and 42CAP, we had the pleasure of speaking again with Andre about their open-source strategy, product differentiation, and scaling their team.

How Did You Get From Open Source to an Enterprise Product?

Since ChatGPT launched, everyone asked how they could use large language models to process their own data. Turns out that since large language models naturally deal with text in the form of vectors, everyone needs a vector store, and vector databases became a core part of the new AI stack – exactly what we are building.

We have developed an open-source vector database that people can deploy and manage themselves for free – we don’t have any commercial features in the core of the engine and don’t intend to have any. It would be essentially like building two different products. Instead, we’re offering a fully managed cloud service, which covers everything you need to run and maintain the vector database, such as backups, logs, or scaling on demand. 

The open-source approach is great for getting feedback from the developer community, creating trust, and letting developers explore how they could use Qdrant. Yet, for production use, our managed solutions get customers up and running quickly and save them the hassle and overhead of deploying and managing the database themselves.

We started by offering a cloud service and partnering with major cloud providers such as AWS or Azure, where developers can sign up to use Qdrant themselves. Yet, especially larger enterprises prefer an on-premise solution – and they have an entire catalog of questions around performance, compliance, and security. While the cloud solution grows product-led, ticking all the boxes for enterprise customers requires quite some sales and development effort. They even asked about the number of versions we had previously.

Initially, it also took some handholding to win enterprise clients – doing things that we didn’t consider core to our product, but that were necessary to get them started. Machine learning engineers are oftentimes not as familiar with handling databases, and setting them up not correctly might lead to things not working or just very slowly. For example, they may forget to create a database index – a data structure that improves the speed of data retrieval. It’s like sorting a list of strings alphabetically but for vector data. When you forget to create such an index, searching through the data still works, but it may take a long time, especially when the database has billions of entries.

How Do You Differentiate Your Product?

Vector databases have become an integral part of the AI stack. That’s why developers may wonder why there is yet another vector database when they hear about Qdrant for the first time. But the point is that not everything that looks like a vector database performs like a vector database and offers the same features.

Many traditional databases have rebranded themselves as vector databases, but there’s a fundamental difference. They have developed features for processing vector data only marginally so as not to lose customers, while their core architecture is geared towards other use cases. There are only very few competitors building genuine vector databases, and all of them were around even before generative AI became a thing. 

As we’re building open-source, we can’t and don’t want to hide anything. It’s always possible for others to copy and paste our code and build on it, which is not a bad thing at all. Sometimes, we publish something, and others build on it, and sometimes, they publish something, and we integrate it.

What makes Qdrant special is how we build our open-source vector database and the surrounding tools – our tech stack and software architecture, using the programming language Rust to write highly optimized code, which gives us superpowers for building performant and scalable software. 

Performance is hugely important for us, e.g., to address real-time use cases. For example, a self-driving car needs to recognize road signs and, therefore, might request from the database the ten most similar road signs to the one its camera recorded. Other databases may run garbage collection first, a process for finding and deleting data entries that are no longer referenced by other entries, which costs valuable time. This can’t happen with our database, as it is optimized for performance and working with large data.

In addition, we provide an amazing developer experience, which makes it really easy for people to get started using Qdrant – depending on your needs, we offer a transparent API, a Docker clone, a Python library, and easy ways to deploy Qdrant on your server or try out our cloud solutions. Simplicity is a major decision factor: Other solutions may be scalable too and benchmark their performance, but they’re super complex to deploy and manage.

Where Will Qdrant Be in a Few Years?

With our latest funding round, we’re in a position to build Qdrant into the best solution for scalable vector search and become an integral part of the AI stack over the next years. 

We spent a lot of time building the product, and we’re now experiencing more inbound interest than we could handle – so it’s time to scale up. We currently have a team of around 40 people, most of whom are engineers, and sales have been founder-led, as it should be in an early-stage startup. We will build a sales organization while continuing our product development to realize our vision for Qdrant.

Comments are closed.