Here’s why vector databases are a big deal for AI

  • Vector databases and capabilities aren't new but are hot right now thanks to interest in GenAI

  • Vectors allow contextual metadata to be encoded numerically into images and other data used to train AI models

  • A battle is brewing between different kinds of vector vendors as adoption is expected to rise dramatically in the coming years

Shiny new toy or knight in shining armor? It seems like vector database capabilities could be the latter for enterprise customers interested in adopting advanced generative artificial intelligence (GenAI).

“Although vector databases predate the adoption of generative AI in the enterprise, their ability to enable rapid scalability and performance of generative AI applications has resulted in an explosion of their popularity and adoption during the past 12 months,” Gartner Analyst Arun Chandrasekaran told Fierce.

But hold on one second. What even is a vector much less a vector database? We took the question to cloud database giant MongoDB. Like many others in the industry, MongoDB has introduced vector capabilities for its Atlas platform.

What is a vector database?

As MongoDB SVP of Product Andrew Davidson explained it, vectors are just lists of numbers that encode meaning into pieces of structured or unstructured data. Critically, vectors don’t just convey what an object is, but also key context about it. Or, as Futurum Group Research Director Keith Kirkpatrick put it, vectors help computers and AI models capture the “essence” of a piece of data.

If you took a psychology class in college, you probably learned about schemas or object schemas, which are frameworks humans use to classify and organize information about the world. Vector databases are kind of like that, but, you know, for computers.

Here’s an example of how it works. Head back to grade school for a moment and imagine a very simple chart with an X axis and a Y axis. Now, picture a few points on it labeled pen, pencil, notebook and paper. Pen and pencil are close together as are notebook and paper.

“While a computer does not know that a pen and a pencil are related to each other or more similar to each other than a pen and a notebook, we can encode it in that number space…[and] the computer can now immediately deduce that a pen is close to a pencil,” Davidson said. “So, you can now ask the computer ‘give me a nearest response quickly, give me related things.’”

And having that sort of building block is a big deal when you put it in the context of generative AI.

Kirkpatrick explained that generative AI use cases are starting to move beyond basic text-based inference tasks and on to more complex tasks like text-to-image or text-to-video generation. In order to complete those tasks, AI needs to “understand the essence of things, not just keyword matches,” he said. But individually inputting every piece of data needed to enable that would be massively time consuming.

That’s where vector databases – and the encoded metadata they contain – come into play. They offer a way for companies to “provide long-term memory for stateless generative AI models and can boost their accuracy and reduce their hallucinations through prompt augmentation,” Gartner’s Chandrasekaran said.

All aboard!

At this point, most of the big names out there have rolled out vector capabilities. Think AWS, Microsoft, IBM, Databricks, MongoDB, Salesforce, and Adobe. There’s also a raft of vector database specialists in play, including Chroma, Qdrant (which just debuted a version of its offering in a hybrid cloud model last month) and Pinecone.

But there are some interesting divides emerging in the vector arena: one between open- and closed-source players and another between dedicated vector databases and databases that offer vector storage and search capabilities. On dedicated, open-source side you have the likes of Chroma, Qdrant and Milvus (which works with IBM), while Pinecone is a leading dedicated, closed-source player. Snowflake, meanwhile is open source and offer vector search capabilities but isn’t necessarily a dedicated vector database.

And there’s good reason for so many to be jumping into this arena. Gartner has predicted 30% of enterprises will use vector databases to ground their generative AI models by 2026, up from 2% in 2023.

“There is a looming battle for customer wallet share between closed-source and open-source vector databases, as well as between purpose-built vector databases versus incumbent databases and search vendors adding vector storage and retrieval as an add-on capability,” Chandrasekaran said.

Risky business?

Gartner recently flagged raw data leakage from embedded vectors as a potential risk associated with using vector databases for Generative AI.

Given compute costs associated with AI, Chandrasekaran said it would also be wise for enterprises to invest in workforce training around vector database capabilities “to avoid expensive misuse and misalignment.”

But according to Kirkpatrick the biggest risk is “assuming that the technology will just work. That is the biggest potential pitfall with this.”

Without oversight, there’s the potential for the vector-fed AI models to drift toward one erroneous bias or another. The key to success, he concluded, will be making sure the models is “returning what it should” in response to prompts.”