Vector Search Inside Out

Every "find similar" feature, from semantic search to the retrieval step in a chatbot, ends in the same question: which of my millions of stored vectors are closest to this one? HNSW is the algorithm almost every vector database uses to answer it in a fraction of a millisecond. Most of the engineers who depend on it have never seen inside it.

This book opens it up. You follow Asha, a forward deployed engineer, as she builds "find similar" products and chatbot retrieval for a retailer with 1.2 million product embeddings. Each idea arrives at the moment she needs it, explained in plain English and shown in 92 step-by-step figures drawn from animated explorables.

Part One builds HNSW from nothing:

- what "nearest" means, what recall measures, and when exact search is simply the better tool

- how a search walks a graph, why a greedy walk gets stuck, and how a beam of width ef rescues it

- why the index has layers, how a query descends them, and how each point chooses its links

- how the index builds itself, one search per point

Part Two turns every knob on real data, 60,000 Fashion-MNIST images and 1.18 million GloVe word vectors:

- M, efConstruction and ef: what each controls, and what each costs in recall, latency and memory

- why fixed settings lose recall as the collection grows, and a five-step tuning recipe for your own data

- memory, quantisation and re-ranking: a PQ index 13× smaller, with recall pulled back from 0.82 to 0.997

- filters, deletes and updates, which distance to use, what nine databases call each knob, and when HNSW is the wrong tool

You need basic Python and a rough idea of what an embedding is. No maths beyond distances, no prior knowledge of graphs. Every number was measured with hnswlib and faiss, and every chapter builds on the one before.