HNSW vs IVF: how do vector indexes trade off recall, speed and memory?

Exact nearest-neighbor search compares the query with every vector, which is too slow at scale, so we use approximate indexes.

HNSW builds a multi-layer graph where each vector links to its near neighbors. A search starts in a sparse top layer and walks greedily toward the query, dropping to denser layers. It gives excellent recall at low latency and handles inserts well, but the graph links add a lot of memory on top of the vectors, and builds are slower. Raising efSearch improves recall at the cost of latency.

IVF first clusters the vectors, for example into a few thousand lists, and at query time searches only the nprobe closest lists. It uses less memory, and with product quantization the vectors themselves are compressed, which lets billions of vectors fit. The cost is lower recall when the answer sits in a cluster you did not probe, and clusters need retraining as data shifts.

I would use HNSW when the index fits in memory and latency matters most, and IVF-PQ when memory or scale is the constraint. Either way I measure recall@10 against exact search on a sample of queries before choosing parameters.