What Is Cosine Similarity?

How Cosine Similarity Works

Cosine similarity answers a narrow question: do these two vectors point in the same direction? It takes the angle between two vectors and returns its cosine. When the angle is zero — the vectors point exactly the same way — the cosine is 1. When they are perpendicular, meaning unrelated, it is 0. When they point in opposite directions, it is -1. Because only the angle matters, the length of each vector is irrelevant, which is the property that makes cosine similarity well suited to comparing meaning.

That property matters for AI search because a vector embedding encodes meaning as direction in a high-dimensional space. A short passage and a long one about the same subject may have very different vector lengths but nearly the same direction, and cosine similarity correctly scores them as highly similar. This is why it is the default metric for comparing embeddings and for measuring semantic similarity between a query and a candidate passage.

Formula

Cosine similarity of two vectors A and B is their dot product divided by the product of their magnitudes:

                    A · B              Σ (Aᵢ × Bᵢ)
cos(θ)  =  ───────────────────  =  ─────────────────────────
                 ‖A‖ × ‖B‖          √(Σ Aᵢ²) × √(Σ Bᵢ²)

The numerator multiplies the vectors component by component and sums the results. The denominator divides by each vector’s length, which is what strips out magnitude and leaves pure direction. If both vectors are already normalized to length 1, the denominator equals 1 and the whole formula collapses to just the dot product A · B.

Example of Cosine Similarity

OpenAI’s embeddings documentation is a clean worked example of cosine similarity in production. The docs explicitly recommend it: “We recommend cosine similarity” as the way to measure how related two embedding vectors are. They also note a practical shortcut that follows directly from the formula. Because OpenAI embeddings are normalized to length 1, the denominator in the equation is already 1, so cosine similarity can be computed “slightly faster using just a dot product,” and cosine similarity and Euclidean distance “will result in the identical rankings.”

Sentence-BERT shows the same metric doing the heavy lifting in retrieval. In the 2019 SBERT paper, each passage is embedded once, and comparing 10,000 of those embeddings to find the most similar pair takes about 0.01 seconds — because the comparison is nothing more than cosine similarity between precomputed vectors. That is the whole point: embeddings make meaning comparable, and cosine similarity is the cheap arithmetic that does the comparing. Every time an AI answer engine decides which passage best matches a question, a cosine score is what ranks the candidates.

Frequently Asked Questions

The Bottom Line

Cosine similarity reduces "how alike are these two things" to "how small is the angle between their vectors," scored from -1 to 1. It is the arithmetic that turns embeddings into retrieval: a query vector is compared against every passage vector, and the closest directions win. What matters is not the raw score but where your passage ranks against the others competing for the same query.