The text discusses the pivot from exact to approximate search results in Couchbase's Search paradigm. The authors introduce vector search (KNN) with its distance-based similarity scoring, which necessitates a shift in how relevance is measured. They explore the trade-off between recall and latency in KNN searches, particularly when using FAISS for vector index creation, training, searching, and related functionality. Through experimentation and tuning, they find that increasing centroids can decrease recall, while increasing nprobe (number of nearest neighbors to consider) improves recall but increases search time. They propose a solution by introducing a dynamic nlist (number of points per cluster) formula based on the dataset size, allowing users to optimize for either recall or latency. This pivot enables Couchbase Search to provide more accurate and relevant results while maintaining flexibility in user preferences.