Cosine similarity and dot product give identical rankings under exactly one condition. Which is it?

Cosine similarity between two vectors u and v is defined as the dot product divided by the product of their L2 norms: `cos(u, v) = (u · v) / (||u|| · ||v||)` The numerator is the dot product. The denominator is the product of the lengths. **When every vector has unit length** (`||u|| = ||v|| = 1`), the denominator is 1 and cosine equals dot product. Same value, same ranking. **When vectors have different lengths**, the dot product gets biased: a longer vector that points roughly toward the query can score higher than a shorter vector that points exactly at the query. Cosine ignores length entirely. This matters in vector DBs because dot product is computationally cheaper than cosine (no square root, no division). Most modern embedding models (OpenAI text-embedding-3-small/large, BGE, E5) return L2-normalized vectors as a contract. That means you can configure the index to use dot product and get cosine ranking semantics for free. The other answer options (mean-centering, dimension parity, index family) are unrelated to the equivalence question.

**L2 normalization is the only condition** under which cosine and dot product give identical rankings. The math is direct. Cosine similarity is: $$\text{cos}(u, v) = \frac{u \cdot v}{\lVert u \rVert \, \lVert v \rVert}$$ If `||u|| = ||v|| = 1` for every vector in the corpus and for every query, the denominator collapses to 1 and the expression reduces to the dot product. Rankings are identical because the two functions are now literally the same function. Without normalization, the denominator varies per pair, and dot product picks up a length bias: a vector with magnitude 2 pointing in a slightly off direction can outscore a vector with magnitude 1 pointing precisely at the query. Cosine eliminates this by construction. **The 2026 production pattern** exploits the equivalence. OpenAI text-embedding-3-small, text-embedding-3-large, Cohere v3, BGE, and most sentence-transformers return L2-normalized vectors as part of their contract. You configure the vector DB with dot-product distance (cheaper, hits SIMD-optimized kernels directly), and you get cosine ranking semantics with zero overhead. The distractor options (mean-centering changes covariance structure but not norms, dimension parity is irrelevant, the HNSW versus IVF distinction is about index family not metric semantics) all conflate orthogonal concerns.