Vector Index

When a vector database receives a query vector, it must identify the K vectors in a potentially billion-item corpus that are most similar (by cosine similarity or Euclidean distance). Comparing the query against every stored vector — exact nearest-neighbor search — is prohibitively expensive at scale. Vector indexes solve this by organizing the vector space into a structure that allows large fractions of the corpus to be skipped during search, returning approximate results with controllable recall loss.

The dominant index algorithms are Hierarchical Navigable Small World (HNSW), which builds a layered graph where higher layers provide coarse navigation and lower layers provide fine-grained proximity links; Inverted File Index (IVF), which clusters vectors into Voronoi cells and searches only the nearest cells; and Product Quantization (PQ), which compresses vectors into compact codes to reduce memory footprint. Real-world deployments frequently combine these — IVFPQ is a common pairing — to balance RAM usage, build time, query latency, and recall.

Enterprise architects must understand that index parameters are not one-size-fits-all. The HNSW parameters "M" (number of bidirectional links per node) and "ef_construction" (search depth during build) directly govern the recall-latency-memory triangle. Larger values improve recall but increase build time and RAM. Additionally, some workloads require filtered ANN search — finding the nearest neighbors among vectors that also satisfy a structured metadata predicate — which many naive index implementations handle poorly. Purpose-built vector databases have invested heavily in efficient pre-filtering and post-filtering strategies that enterprise teams should evaluate carefully.