HNSW vs IVFFlat: How to Choose the Right Vector Index

HNSW and IVFFlat are the two dominant ANN vector index strategies. HNSW is a hierarchical graph index offering 95%+ recall out of the box, handles continuous inserts without rebuilds, but uses 2-5x more memory. IVFFlat partitions vectors into k-means cells, builds faster, uses ~1.1x raw vector memory, but recall drifts as data changes and requires periodic rebuilds. HNSW is the recommended default for most workloads under ~10M vectors with active writes; IVFFlat suits very large (50M+), mostly static datasets where memory or build time is the bottleneck. The post covers how each algorithm works internally, parameter tuning (m, ef_construction, ef_search for HNSW; lists and probes for IVFFlat), concrete benchmark numbers using pgvector, a decision framework with crossover thresholds, and a look at emerging alternatives like binary quantization, IVF+RaBitQ, DiskANN, and ScaNN.