Vector search explained: encode, normalize, compare, rank—and why ANN wins at scale

techNmak · x · 2026-10-02

A clear walkthrough of embedding-based vector search, broken into four operations: encode, normalize, compare, rank.

Documents pass through an embedding model (all-MiniLM-L6-v2, 384-dim in the example); token reps are mean-pooled excluding padding, L2-normalized, and stored alongside passage IDs.
Queries go through the same pipeline, so queries and docs share one embedding space.
With unit-norm vectors, cosine similarity reduces to a dot product: score, sort, return top-k.
At scale, exact scan cost grows linearly—one query over 1M×384 vectors means 384M coordinate multiplications—so systems use approximate nearest-neighbor indexes (HNSW, IVF) that trade some exactness for speed.
Key caveat: vector similarity measures embedding-space closeness; it doesn't prove a retrieved passage is correct, complete, or relevant.

More from Research