This blog is the second post in a series on building a billion-scale vector search using Vespa. The series covers the cost and serving performance tradeoffs related to approximate nearest neighbor search at a large scale. In the first post, we introduced some challenges related to building large-scale vector search…

How fast is fast? Many consider the blink of an eye, around 100–250ms, to be plenty fast. Try blinking 10 times in a second? If you can manage, your blink-of-an-eye latency is around 100ms. …

Doug Turnbull asks an interesting question on Linkedin; Will new vector databases dislodge traditional search engines? The short answer is no, but it depends on what you classify as a traditional search engine. Features like phrase search, exact search, BM25 ranking, dynamic summaries, and result facets are features we take…

There is a growing interest in AI-powered vector representations of unstructured multimodal data and searching efficiently over these representations. This blog post describes how your organization can unlock the full potential of multimodal AI-powered vector representations using Vespa — the industry-leading open-source big data serving engine. Photo by israel palacio on Unsplash Introduction Deep Learning has revolutionized information extraction from unstructured data like text, audio, image, and videos. Furthermore, self-supervised learning algorithms like data2vec accelerate learning representations of speech, vision, text, and multimodal representations combining these modalities. …

Find videos and links from four Vespa-related talks at Berlin Buzzwords 2022, Germany’s most exciting conference on storing, processing, streaming, and searching large amounts of digital data. — Berlin Buzzwords 2022 has just finished, and we thought it would be great to summarize the event. Berlin Buzzwords is Germany’s most exciting conference on storing, processing, streaming, and searching large amounts of digital data, focusing on open-source software projects.

This blog post describes HNSW-IF, a cost-efficient solution for high-accuracy vector search over billion scale vector datasets. — The first blog post on billion-scale vector search covered methods for compressing real-valued vectors to binary representations and using hamming distance for efficient coarse level search. The second post described approximate nearest neighbor search tradeoffs using Hierarchical Navigable Small World (HNSW), including memory usage, vector indexing performance, and query performance…

This blog post describes Vespa’s industry-leading support for combining approximate nearest neighbor search, or vector search, with query constraints to solve real-world search and recommendation problems at scale. Photo by Christopher Burns on Unsplash This blog post describes Vespa’s support for combining approximate nearest neighbor search, or vector search, with query-time filters or constraints to tackle real-world search and recommendation use cases at scale. The first section covers the data structures Vespa builds to support fast search in…

In the first post in this series, we introduced compact binary-coded vector representations that can reduce the storage and computational complexity of both exact and approximate nearest neighbor search. This second post covers an experiment using a 1-billion binary-coded representation derived from a vector dataset used in the big-ann-benchmark challenge. …

In this blog post, I investigate running Machine Learned (ML) inference computation as close as possible to where the model input data is generated on a user's device at the edge. There are several definitions of "Edge," but for me, Edge computations cover inference computations on the user device, in…