DocsConcepts

Indexing

Exact nearest-neighbor search is O(N·d). Approximate indexes trade a tiny bit of recall for orders of magnitude in speed.

HNSW — Hierarchical Navigable Small World

A multi-layer graph where each node is connected to a handful of neighbors. Queries enter at the top sparse layer and greedily descend, dramatically cutting comparisons. The default choice in Qdrant, Weaviate, Pinecone serverless and pgvector ≥ 0.5.

Build time

Slow

Query latency

Very fast

Recall

0.95–0.99

Memory

High (in-RAM graph)

IVF — Inverted File

Partitions vectors into nlist Voronoi cells via k-means. A query only scans the closest nprobe cells. Combined with product quantization (IVF-PQ) it compresses each vector to a few bytes — perfect for billion-scale indexes that don't fit in RAM.

text

recall  ↑ nprobe ↑  latency ↑
recall  ↓ nprobe ↓  latency ↓

DiskANN & Vamana

Graph indexes optimized for SSD. The graph lives on disk and only the search path is paged in. Used by Milvus 2.x and Microsoft Bing for tens of billions of vectors on a single node.

Pick HNSW unless you have a reason

For most teams < 100M vectors, HNSW with default parameters is the right answer. Reach for IVF-PQ or DiskANN when memory becomes the binding constraint.