VECTOR DATABASES IN DEPTH

Vector Databases in Depth is a practical engineering guide to the data layer that makes semantic search, recommendation, and retrieval-augmented generation practical. It starts from a single observation: vector search is a systems problem, not a library call, and the difference between a prototype and a production system is measured in the same units as any other distributed system — recall, latency, cost, isolation, and operational discipline. The book walks through the full stack — why vector search matters and where it fits, embeddings as a data type and the decisions that shape them, similarity metrics and the normalisation practices that make them interchangeable, approximate nearest-neighbour search and the recall-speed trade-off, HNSW in detail with its parameters and tuning workflows, IVF-PQ and disk-based indexes for corpora that exceed memory, quantization from int8 to binary and the rescoring patterns that recover lost recall, metadata filtering that participates in the index traversal, hybrid retrieval that fuses dense and sparse results, re-ranking and multi-stage pipelines, horizontal scaling with sharding and replication, multi-tenancy with enforced filter isolation, operations, monitoring and cost engineering, and the trends reshaping the field. It covers the failure modes that quietly wreck vector search: an embedding upgrade that invalidates every cached vector, a filter applied after retrieval that excludes the top-k, a shard that becomes hot because the tenant hashing function is not uniform, a quantization step that reduces recall by twenty points without any metric catching it, a re-ranker that adds three seconds of latency, an index that fits at build time and does not fit at query time, a multi-tenant query that leaks vectors across tenants. Each is presented with the failure, the countermeasure, and the operational tradeoff.