This confirm my real-world usage. We had to do a "Semantic Similarity search" for some documents, where we represent a document as a vector using BERT, and had to look for documents close to a reference document.

First we did on top of an ANN library(Facebook FAISS). For simplicity we moved to the naive Vespa vector search and the performance was good enough( FOR 1Mi documents, FAISS did the search in 50ms, Vespa was around 600ms), and we moved to Vespa to reduce the number of tools to maintain.