June 2022 Summaries

The June 2022 Vespa newsletter highlights the release of Vespa 8, which now runs on Java 17 and CentOS Stream 8, offering several new features and improvements. These include support for pre- and post-filtering in Approximate Nearest Neighbor (ANN) searches, allowing users to balance precision and performance effectively, as well as fuzzy query matching using Levenshtein edit distance to improve search accuracy. Vespa also introduces embedding support, enabling the mapping of data into abstract vector spaces for efficient retrieval and computation. Further advancements include fast-rank for reduced latency in tensor ranking expressions and support for importing ONNX models in rank profiles. The newsletter also mentions the availability of new Vespa blog posts and training videos focused on advanced search techniques and system troubleshooting.

Vespa 8, the latest major version of Vespa.ai, has been released for production use both as open source and on Vespa Cloud, but unlike traditional major releases, it introduces no new features. Instead, the release focuses on removing deprecated functionalities, updating default settings, and switching to newer dependencies like JDK and Linux, reflecting a shift in how software updates are managed in the era of continuous deployment and semantic versioning. Users are encouraged to ensure their applications are free from deprecated features and to follow straightforward steps to upgrade, which include setting the major version to '8' in the deployment configuration or switching to CentOS Stream 8 if running Vespa independently. This approach allows for a seamless transition and ensures that users continue to receive new features and improvements promptly while maintaining compatibility across minor releases.

Berlin Buzzwords 2022 showcased a variety of discussions and presentations centered around the processing and searching of large data sets, with a particular focus on open-source software projects. At the event, Vespa.ai's team members delivered talks on significant topics such as AI-powered semantic search and hybrid search models, emphasizing both the potential and challenges of integrating AI with traditional search methods. Kristian Bergum highlighted the limitations of AI-powered semantic searches when applied to new domains, while Lester Solbakken discussed how traditional lexical models and deep retrieval models can complement each other to improve search outcomes. Additionally, Atita Arora presented an insightful comparison between Vespa and Apache Lucene-based search engines, and John Dagdelen showcased Matscholar, a search engine for materials science researchers that utilizes Vespa to enhance search capabilities in scientific research. These talks underscored Vespa's versatility and potential in advancing search and data retrieval technologies, particularly in complex and data-rich environments.

The blog post discusses a cost-efficient hybrid approach for billion-scale vector search using a method called HNSW-IF, which combines Hierarchical Navigable Small World (HNSW) with a disk-backed inverted file system. This method addresses the high memory requirements of in-memory algorithms like HNSW by utilizing solid-state disks to store most vector data while maintaining in-memory graph data structures for efficient search. Inspired by the SPANN approach, the Vespa team implemented this hybrid method, which partitions a vector dataset into clusters, using centroids to represent clusters, and stores non-centroid vectors in disk-based data structures. The Vespa implementation features real-time indexing and querying capabilities and supports CRUD operations, making it suitable for real-world applications with varying query volumes. Experiments show that the hybrid HNSW-IF system can achieve high recall rates with acceptable query latencies, making it a viable solution for applications requiring scalable, accurate vector search without the high costs associated with fully in-memory systems.