As we have noted before, vector databases aren’t new even though people talk about them that way, and in fact can trace their origins back a few decades. Exactly when depends on who you ask, but their popularity began to ramp along with machine learning technologies over the past 10 years and now, with the explosion of innovation and adoption of large language models and generative AI, seemingly everyone wants to be in the vector database business.
In a nutshell, vector is the new JSON.
Major cloud service providers like Amazon Web Services, Microsoft Azure, and Google Cloud are populating their myriad database services with vector search capabilities, and established database makers like Oracle, MongoDB, Snowflake, and DataStax already offer vector capabilities in some of their offerings. There also is a growing list of smaller companies – such as Pinecone, Milvus, Chroma, and Weaviate – that offer open-source vector databases.
Most recently, both Google Cloud and Couchbase expanded vector search capabilities in their database offerings. Google Cloud, which last year announced support for pgvector vector search capabilities on its Cloud SQL for PostgreSQL and AlloyDB for PostgreSQL managed databases, unveiled vector search across more databases, including Spanner, MySQL, and Redis, in conjunction with more integrations with LangChain, a framework for creating language-based applications.
For its part, Couchbase said it was introducing vector search as a feature in its Capella database-as-a-service (DBaaS) and Couchbase Server, enabling enterprises to use vector search capabilities on-premises, in the cloud, and out at the edge in mobile and Internet of Things (IoT) devices. The NoSQL database vendor also is adding support for LangChain and LlamaIndex to aid developers in their use of LLMs.
“The implications for databases are you have to support vector search,” Scott Anderson, senior vice president of product management and business operations at Couchbase, tells The Next Platform. “We were obviously doing that with this announcement, with an idea that it’s part of a larger platform and set of capabilities that application developers require. Because they need different access patterns or they’re writing different types of queries, you need to be able to support all of those: KV for sub-millisecond response time, more complex queries, which can be done just through normal querying, or querying large sets of data in a columnar database. Being able to do full text search, geospatial search, and semantic search are all critical capabilities.”
Generative AI Fueling A Fast-Growing Market
The global vector database market is expected to expand rapidly, from $1.5 billion last year to $4.3 billion by 2028, driven in large part by the expanding use of LLMs and machine learning and the deluge of highly personalized and adaptive AI applications that will be developed and deployed by businesses for search, recommendations, and other use cases.
Organizations can use vector databases to store and access both structured and unstructured data – think text, images, videos, and audio. Vector embeddings typically are generated via machine leaning to give the data objects semantic meanings. Objects that are alike will have similar vectors and are located closer to each other, so searches that use approximate nearest neighbor (ANN) algorithms improve the speed and accuracy of searchers.
“One of the examples that we use is this: I want a pair of blue shoes that match the color of my car,” Anderson says. “I want these brands, Adidas and Nike and ASICS, in this price range, and within 15 miles of my house, and they need to have these couple of sizes and they need to have available inventory. I’m giving you a bunch of inputs and how I process that input is through a broad range. That’s a query on the inventory. It’s geospatial services to understand the stores to be able to find the inventory. Vectorizing the color blue, I upload the photo of my car and then am able to do a vectorized search to see the closest proximity to that color blue that I’m looking for, a range search based off of price.”
Businesses are dealing with consumers who are more specific, populating their searches with more context to get the best results, he says, adding that the example illustrates the impact at the application level. “The database has to be able to provide those capabilities,” he says.
The introduction of vector search is the latest step in Couchbase’s efforts to roll AI capabilities into Capella. That includes Capella iQ, an AI cloud service powered by ChatGPT that developers can interact with via natural language as they write create code and indexes and sample data. At AWS Reinvent, the company rolled out a columnar service for real-time analytics in Capella, which is in private preview.
All Search On A Single Platform
Couchbase is looking to differentiate itself by essentially adding the vector search service to the same platform that delivers other capabilities, rather than offering a single-purpose vector database, Anderson says.
“One of the challenges that you get into with single-purpose vector databases is really the latency, as you’re trying to combine multiple access patterns from an application to be able to return a result,” he says. “Being able to come into a single data platform ensures that your data is consistent across the platform, where the data is all replicated to all nodes. There’s no real latency because we do memory-to-memory replication. If you have a bunch of single-purpose databases, you’re syncing data across those databases and you’re using multiple APIs to be able to access the data vs. being able to access in a single call into a data platform itself.”
Making vector search in all Couchbase products allows for both similarity and hybrid search that combines text, vector, range, and geospatial search, retrieval-augmented generation (RAG) to make AI software more accurate and safer, and a single index for all search patterns to reduce latency.
Mobile: The Next Frontier
A key for database players – including Couchbase – is adapting to the fast-evolving nature of AI, Anderson says. There are new models and other innovations constantly popping up as the AI field unfolds, and vendors like Couchbase need to be “willing to make some bets on what technology you think are going to take off,” while balancing that with the need to keep an open platform and APIs that will allow them to support a broad ecosystem that is constantly changing and growing.
“We’ve always had a full text search, we’ve added geospatial, and a number of the query,” he says. “The effort by our engineering team is bringing in that capability to store vectors efficiently within the JSON documents, the ability to create those indexes and then be able to do that in performance at scale.”
Couchbase also is looking to bring vector search to the edge in mobile and IoT devices, with plans to bring vector search to Couchbase Lite, the vendor’s embedded database. Business users and individuals continue to do more on their smartphones and other mobile devices and more AI applications will be directed that way.
“One of the unique things about our mobile solution is that you can access all the data because we’re persisting it on that device, irrespective of network connectivity,” Anderson says. “It’s very popular with field services, retail and so forth. That unveils more opportunities not just from the cloud, but all the way down to the edge. One of the examples that we’ve talked about is the ability to take a photo with your phone and then have an embedding created off of that and then be able to search on the device to find the closest match to various images. You can imagine that retail settings and many other settings.”
Nice article! I’m glad Couchbase is joining the database vector-search fray, with all of its tentacular trifurcations into the AI/ML space. It’s great (for me) to have learned a new meaning for ANN (beyond Artificial Neural Network), and, from there, get exposed to ANNoy ( https://github.com/spotify/annoy ) and to Machine Unlearning by “Unfolded Self-Reconstruction Locality-Sensitive Hashing” (USR-LSH) in ANN search (and also to “Zero-Shot Machine Unlearning”).
This has made my Saturday! (eh-eh-eh!)