Follow Datanami:
February 29, 2024

Couchbase Bolsters GenAI Development with Vector Search, RAG


The latest release of the Couchbase database adds support for vector search, integration with Llamaindex and LangChain, and support for retrieval-augmented generation (RAG) techniques, all of which are designed to help customers streamline the development of generative AI applications.

Couchbase has gained traction over the years for developing an operational data store that offers a mix of transactional as well as analytic capabilities. With its support for SQL++ (formerly N1QL) as well as the November announcement of the addition of a columnar data type, the JSON store is morphing with customer needs.

With today’s release of Couchbase version 7.6, Couchbase customers are gaining new GenAI capabilities, including vector similarity search, hybrid search, and support for LangChain and LlamaIndex, which are open source tools that help connect customers’ data to large language models (LLMs).

The new vector search capability will leverage vector embeddings stored directly in the JSON document to provide more relevant results for searches. Vector embeddings are mathematical representations of data. In Couchbase’s case, the embeddings could represent facts about a customer, such as their preferences and purchase history. At runtime, the embeddings would be called by the LLM to enhance a query within the Couchbase database.

It’s all about building more personalized experiences within adaptive applications, said Scott Anderson, the senior vice president of product management and business operations at Couchbase.

“[Adaptive applications] are going to…be highly contextualized to be able to deliver a level of hyper personalization, which will allow the end user of that application to get the most relevant information,” Anderson said in a Datanami interview yesterday.


The new hybrid search capability, meanwhile, broadens searches to include multiple data types, including text, vectors, ranges, and geospatial data. That will further help the database to deliver more tailored results for queries.

Couchbase customers have asked for these features, Anderson said. In many cases, customers want to store vector embeddings directly in their operational database for privacy and performance reasons.

“I think getting the embedding models close to the data is important,” he said. “Bringing more of those capabilities close, or co-located with the data platform, is going to reduce complexity and latency, and ensure privacy of the data.”

Couchbase customers today are experimenting with GenAI and in some cases using APIs for LLMs such as the GPT models from OpenAI, Anderson said. However, the company foresees that customers will increasingly want greater control over their GenAI infrastructure, partiucarly when it involves sensitive data.

“You can imagine a financial institution, a healthcare institution, or specific verticals where that data is incredibly valuable for them, and they want to bring the power of AI to that data, which is bringing embedding models within the control of their environment versus potentially accessing out to something where they may lose control of that underlying data,” Anderson said.

As GenAI technology progresses, LLMs are likely to get both smaller as well as more vertically focused, Anderson said. The new GenAI capabilities that Couchbase is adding help prepare the company and its customers to be in a position to take advantage of those technology improvements when they happen, he said.

“I have a hypothesis that [AI models] will get more specific in terms of a vertical industry… [and be] a size that you can run it in the same environment in which your database or data platform resides,” he said. “So potentially within the same VPC…which would reduce latency from a networking perspective.”

The new GenAI capabilities are generally available now in the company’s database-as-a-service (DBaaS) offering, Couchbase Capella, as well as Couchbase Server, which runs wherever a customer wants. The company will be discussing Couchbase version 7.6 in a series of upcoming talks next month.

Related Items:

The Future of Databases Is Now

Couchbase Advances Case for Becoming Your System of Record

Couchbase Pops in Stock Market Debut