Vector Search

Approximate nearest-neighbor (ANN) retrieval over high-dimensional embeddings — the engine behind semantic search, recommendation, and visual similarity.

Vector search indexes a numeric embedding for every item, then answers “find the K items closest to this query embedding.” At scale it uses approximate algorithms (HNSW, IVF, ScaNN) that trade tiny recall losses for 100–1000× speedups over exact nearest-neighbor.

For ecommerce, vector search is the retrieval layer for: semantic search (text query → product), visual search (image query → product), related products (product → product), and personalization (user embedding → product). The same index can serve all four.

Operational concerns: embedding drift when you upgrade the model (you must re-embed the entire catalog), index rebuild cost on catalog changes, and the recall-vs-latency tradeoff in the ANN parameters. Most teams target ≥ 95% recall@10 against an exact baseline.

Related terms