A Short Guide to Tweaking Vespa's ANN Parameters

Jan Böker's blog post provides insights into optimizing Vespa's Approximate Nearest Neighbor (ANN) search parameters, particularly with the introduction of ACORN-1 in the Hierarchical Navigable Small World (HNSW) algorithm. The post discusses different strategies Vespa employs for ANN searches with filters, including pre-filtering HNSW, post-filtering, and the newly introduced ACORN-1 strategy. It details how Vespa automatically chooses a search strategy based on filter hit ratios and how adjusting parameters like filter-first-threshold and approximate-threshold can influence performance, response time, and recall. The blog emphasizes the importance of tweaking these parameters to balance response time and recall, noting that changes in parameters can lead to different outcomes based on the dataset's characteristics. The discussion also touches on using parameters like exploration-slack and exploreAdditionalHits to fine-tune recall and response time further, highlighting the trade-offs between increased recall and response time.