Elastic has revealed significant upgrades to its vector search capabilities through the introduction of the ACORN algorithm and Better Binary Quantization (BBQ) as the standard for dense vectors.
The new ACORN algorithm has been developed to accelerate filtered vector search, with internal benchmarks indicating up to a fivefold increase in speed for filtered queries over high-dimensional datasets. According to Elastic, this improvement targets key areas in enterprise artificial intelligence infrastructure, where the speed of information retrieval is crucial for the performance of various AI-driven applications.
Better Binary Quantization, also known as BBQ, is now the default quantisation method for dense vectors containing 384 or more dimensions in the latest version of Elasticsearch. With this change, Elastic reports not only improved ranking accuracy but also a marked reduction in memory consumption and infrastructure costs, addressing practical deployment challenges for organisations seeking large-scale AI solutions.
Company perspective
"We're committed to giving developers the best tools to build and iterate AI applications at scale," said Ajay Nair, General Manager, Platform at Elastic. "ACORN for filtered vector queries and default Better Binary Quantization represent a step-change in performance and efficiency. This enables our users to execute complex, high-speed, filtered queries at low latency with a dramatic memory reduction, all while maintaining high ranking quality."
ACORN-1 is described as a new algorithm for filtered k-Nearest Neighbour (kNN) search within Elasticsearch. It works by directly integrating filtering into the HNSW (Hierarchical Navigable Small World) graph traversal, which underpins Elastic's approximate nearest neighbour search engine. Previous industry techniques typically processed filters after an initial query or prior to indexing, restricting flexibility for developers who wish to apply or change filters on the fly post-ingestion.
The new approach taken by ACORN allows users to define filters at the time of the query, even after relevant documents are in the database. This is particularly relevant for companies handling extensive, dynamic datasets, where filtering criteria may frequently change.
In benchmark tests based on real-world filtered vector searches, Elastic claims that ACORN achieved up to a 5X reduction in query latency while maintaining precision in results. This combination aims to meet the need for both speed and accuracy in business-critical AI search processes.
Advancements in ranking with BBQ
The introduction of BBQ as the default method for quantising high-dimensional vectors marks a notable adjustment in Elastic's search infrastructure. By focusing on aggressive binary compression - approximately a 32-fold reduction compared to previous formats - BBQ lowers the demands on memory and computing resources. This can result in lower infrastructure costs for organisations operating at scale.
When tested on ten benchmarks using the BEIR datasets and evaluated with the NDCG@10 metric, which measures the quality of top-10 rankings, BBQ surpassed traditional float32-based search methods in nine out of ten cases. Elastic states that this improved efficiency is achieved without sacrificing accuracy, due to the algorithm's capability to examine a larger pool of candidates while retaining compact memory use.
These changes are now implemented in Elasticsearch 9.1, targeting developers and data scientists looking for more effective ways to handle AI workloads that require scalable, low-latency search and ranking.
Elastic's upgrades come at a time when many enterprises are exploring ways to make AI models more cost-efficient while ensuring dependable performance. By enhancing both the speed of query processing and the efficiency of infrastructure use, these features contribute tools for organisations to better support large language models, recommendation systems, and other vector-based AI applications.
Further technical details about the deployment of ACORN and BBQ in Elastic's environment are available through company documentation and blog posts, focusing on practical steps and use-case guidance for developers and technology teams.
