HNSW vs IVF, when do you pick each for a production vector index?

Vector databases store millions of dense embeddings and answer 'find the K closest to this query' in milliseconds. They achieve that with approximate nearest neighbor (**ANN**) indexes. HNSW and IVF are the two dominant families, and they trade memory against speed differently. **HNSW** builds a graph where each vector is connected to a few of its nearest neighbors in a hierarchical structure. To search, you start at a random entry point and greedily hop toward the query along edges. The hierarchy means most queries finish in a few dozen hops, giving sub millisecond latency at high recall. The cost is memory: storing the graph adds 2 to 4 times the raw vector size. **IVF** (Inverted File) partitions vectors into clusters using k-means, then at query time only searches the closest few clusters. Memory overhead is small, but search latency is higher because you scan more candidates. Combined with **Product Quantization** as `IVF-PQ`, you compress each vector by 10 to 50 times, making billion scale indexes affordable in RAM.

**HNSW when latency and recall matter; IVF (usually IVF-PQ) when scale or memory is the binding cost.** HNSW is a graph based index. Each vector lives in a hierarchical small-world graph with logarithmic search depth. It hits sub millisecond latency at recall above 0.95 on corpora up to about 100 million vectors. The catch is memory. Storing the graph itself adds roughly 2 to 4 times the raw vector size. A 100M vector index in 1024 dimensions can easily need hundreds of gigabytes of RAM. IVF is partition based. Vectors are clustered with k-means, and search probes only the nearest few clusters. Recall is lower than HNSW at the same compute budget, but the index itself is light. The real win is composing it with Product Quantization (`IVF-PQ`), which shrinks each vector to a few bytes. A billion vector index that would cost a fortune as HNSW can fit on a single machine as IVF-PQ. A practical decision rule: under 10M vectors, HNSW. From 10M to 100M, HNSW if quality first, IVF-PQ if cost first. Above 100M, IVF-PQ almost always.