Fast, Approximate Vector Queries on Very Large Unstructured Datasets

Explore a groundbreaking approach to processing vector queries on massive unstructured datasets in this 14-minute conference talk from NSDI '23. Discover Auncel, a novel vector query engine that offers bounded query errors and latencies for applications with strict service level objectives. Learn how the system exploits local geometric properties of individual query vectors to build precise error-latency profiles, enabling efficient sampling and processing of data while meeting error and latency requirements. Examine the distributed solution's scalability and performance, with experimental results showcasing up to 10x improvement in query latency compared to state-of-the-art approximate solutions. Gain insights into Auncel's ability to process vector queries on the DEEP1B dataset, containing one billion items, in just 25 ms using four c5.metal EC2 instances.