3 Minutes to Start Your Research in Nearest Neighbor Search (romanbikbulatov.bearblog.dev)

A short primer titled "3 Minutes to Start Your Research in Nearest Neighbor Search" walks through the fundamentals, uses, and evaluation of nearest neighbor (NN) methods — the backbone of recommendation systems, image retrieval, face matching and molecule similarity. The author emphasizes that modern systems encode items as vectors in high-dimensional spaces (dozens to hundreds of dimensions) and that NN search is typically framed as either exact (guaranteed nearest) or approximate (faster, "good enough"), with approximate methods prevailing in recommender settings where latency and diversity matter more than absolute precision. Technically, the note highlights the common strategy across algorithms: avoid checking every point by partitioning or navigating the space. It compares brute-force, ball trees (spatial partitioning by “balls”), and HNSW (multi-layer proximity graphs that quickly zoom in then refine locally). Data are simple numeric vectors (e.g., SIFT1M: 1M points at 128 dims with 10k query ground-truth) used to benchmark recall@k, query latency and build time, plus deeper metrics like memory access cost. The piece closes with practical research directions — distance metrics, scaling to billions of vectors, hybrid algorithms, and preprocessing — useful starting points for anyone entering NN search research.