The original title is "Vector Search — how HNSW finds nearest neighbours"

Short summary

HNSW (Hierarchical Navigable Small World) powers vector search in FAISS, pgvector, Qdrant, Weaviate, and Milvus by building a multi-layer graph where search is a greedy walk from sparse long-range links to dense local links. It reduces nearest-neighbor search from millions of comparisons to roughly 1,800 hops, cutting latency from over a second to ~2ms. The layered skip-list-like structure enables efficient approximate search that usually finds the true global nearest neighbor.

• •HNSW builds a multi-layer graph enabling greedy-walk search across vector embeddings
• •Reduces 2M-document search from ~1,240ms brute-force to ~2ms via ~1,800 comparisons
• •Powers major vector databases: FAISS, pgvector, Qdrant, Weaviate, Milvus

Generated with AI, which can make mistakes.