Define L2 normalization for an embedding vector

**L2 normalization** rescales a vector so its length becomes 1. The formula is: $$v_{\text{norm}} = \frac{v}{\|v\|}$$ where `||v||` is the L2 norm, the square root of the sum of squared components. The direction of the vector stays the same. Only the length changes. After normalization the vector lies on the **unit sphere**: the surface of all vectors with length 1. Why bother? Two reasons. - On unit-length vectors, the **dot product equals cosine similarity**. Search becomes a single matrix multiply instead of cosine with extra divisions. - All vectors live on the same scale, so distances and similarities become directly comparable. Almost every modern embedding API returns normalized vectors by default. Production retrieval stacks rely on the invariant for both speed and consistency.

**L2 normalization is dividing a vector by its L2 norm** to produce a unit-length vector pointing in the same direction: $$\hat{v} = \frac{v}{\|v\|_2}, \quad \|v\|_2 = \sqrt{\sum_i v_i^2}$$ The geometric effect is to project the vector onto the unit sphere. The angular relationship between any two vectors is preserved. **Why every production stack does this.** On unit vectors, dot product equals cosine similarity. That collapses two metrics into one and lets the ANN engine use its fastest code path (inner product) without any quality cost. Pinecone, Weaviate, Qdrant, pgvector all run faster on normalized data, and most managed providers auto-detect and route accordingly. **The boundary case.** L2 normalization fails on the zero vector: division by zero. Any production pipeline needs a check that `||v|| > 0` before dividing. OpenAI text-embedding-3, Voyage v3, Cohere embed v3, and BGE-M3 all return normalized vectors out of the box, but self-hosted models often do not. The combination of 'normalize at index and query' plus 'use dot product as the metric' is the default playbook in 2026.