Filtered ANN Search With Composite Vector Indexes

This blog post is part of a series exploring composite vector indexing in Couchbase, focusing on their significance, implementation, and performance. It uses a Smart Grocery Recommendation System as an example to illustrate how composite vector indexes are constructed using FAISS index factory strings for efficient indexing and querying. The process involves embedding relevant text fields into semantic vectors through a transformer model, which are then stored alongside product data for accurate Approximate Nearest Neighbor (ANN) searches. Couchbase's approach ensures scalable vector searches by integrating scalar filtering and continuous updates, moving beyond standalone FAISS indexes. The post also delves into creating and building vector indexes, emphasizing the importance of a sufficient number of documents for training to achieve effective results. It describes the scan process for vector queries, which involves scalar filtering, vector distance computations, and streaming results back to the client, highlighting the role of scan parallelism, scalar selectivity, and pagination. The flexibility in index definition allows tailoring to specific workloads, enhancing query performance by optimizing the order of index keys for different pruning strategies. The post concludes with a look ahead to the next installment, which will explore filtered ANN searches with composite vector indexes, aiming to combine distance-based similarity with application-specific ordering efficiently.