Return of Redis creator bears fruit with vector set data type

LLM query caching also lands soon

The return of Redis creator Salvatore Sanfilippo has borne fruit in the form of a new data type - vector sets - for the widely used cache-turned-multi-model database.

Inspired by sorted sets, one of Redis' core data types known for handling ordered collections, vector sets offer a native way to store and query high-dimensional embeddings, with a focus on vector similarity search for AI workloads.

Redis started life in 2009 as an attempt to build a performant key-value database. By late 2020, it was the most popular database on AWS, thanks to its popularity as a cache and message broker in cloud-native application stacks. Redis has since broadened its ambitions, adding features for machine learning and support for JSON documents in a bid to evolve beyond its caching roots.

Sanfilippo, better known by the nickname antirez, stepped down as the maintainer of Redis in mid-2020, saying he wanted to focus on writing code rather than managing the project.

He returned in December last year to become “a bridge between the company and the community, but also somebody that could produce programming demos, invent and describe new patterns, write documentation, videos and blog posts about new and old stuff,” according to his blog, referring to Redis the company that continues to steer development of the software.

The new data type is the result of Sanfilippo's return, and builds on Redis's sorted sets by enabling storage and querying of high-dimensional vector embeddings - commonly used in generative AI apps to represent semantic meaning in LLMs, for example.

Redis already supports vector similarity search through its Query Engine, introduced in 2023, which allows developers to find vectors most similar to a target input based on proximity metrics like cosine similarity.

Rowan Trollope, CEO of Redis, told The Register that vector sets expose a lower-level API, giving developers more direct control over the underlying vector data - a shift aimed at flexibility and performance.

“It’s extremely fast, easy to compose, and flexible in terms of how you implement it in your application. It's very much in keeping with the Redis ethos. Sanfilippo had to rewrite the entire underpinnings of the vector database which is called HNSW - hierarchical, navigable, small worlds - as a new algorithm,” he said.

Vector sets land in beta form as part of the Redis 8 Community Edition in May, with features and APIs subject to change based on feedback.

Redis has also announced LangCache, which is described as a semantic caching service for AI apps and agents. The goal is to reduce expensive and latency-prone calls to LLMs by caching previous responses to semantically similar queries.

LangCache is set to launch as a fully managed REST service that sits between the application and the LLM, intercepting requests to serve cached responses when possible.

“It's not as straightforward as caching in the database world when you have a deterministic query, like a SQL string. In the case of a language model, you need a probabilistic or semantic cache, and so we use our vector database underneath the covers for that with a fine-tuned embedding model that's optimized for caching,” Trollope told us. ®