Complete the identity: when do dot product and cosine coincide?

The cosine formula is: $$\cos(u, v) = \frac{u \cdot v}{\|u\| \, \|v\|}$$ The dot product `u · v` is the numerator. The denominator is the product of the two vector lengths. **When both vectors have length 1**, the denominator becomes 1 × 1 = 1 and the whole formula collapses to just the dot product. That is the condition: both vectors must be **L2-normalized** (also called unit-norm, unit-length, or simply normalized). This is why production embedding pipelines normalize their vectors. Once normalized, you can search with the cheaper dot-product operation instead of cosine and get the exact same results. The cost is the one-time normalization step at index and query time.

**Dot product equals cosine exactly when both vectors are L2-normalized**, that is when `||u|| = ||v|| = 1`. The reason is one line of algebra: $$\cos(u, v) = \frac{u \cdot v}{\|u\| \, \|v\|} = \frac{u \cdot v}{1 \cdot 1} = u \cdot v$$ This identity is the foundation of the normalize and use dot pattern that almost every 2026 retrieval stack runs. Cosine requires two extra square-root operations per pair (one for each norm). Dot product is a single sum of element-wise products. On modern hardware that means 30-50% lower latency for the same ranking quality. The condition is BOTH vectors must be unit norm. A common confusion is to call a vector 'unit' just because it lives in a high-dimensional space; the unit-norm property is about magnitude, not dimensionality. OpenAI text-embedding-3, Cohere embed v3, BGE-M3, and Voyage v3 all return L2-normalized vectors by default for exactly this reason.