85 agentic tools audited for security incidents · Talk to the lab →
MotionticBook a call
Glossary

What is cosine similarity?

Cosine similarity is a measure of how closely two vectors point in the same direction, computed as their dot product divided by the product of their lengths, ranging from -1 to 1 and ignoring magnitude.

Checked

Embedding models place text as points in a high-dimensional space where similar meanings point in similar directions. Cosine similarity measures the angle between two of those vectors: 1 means the same direction, 0 means unrelated, -1 means opposite.

It is the default distance for most text embeddings and the metric behind most RAG retrieval and semantic caches. It is also widely misread.

Cosine, dot product and Euclidean distance

For vectors scaled to length 1, the three metrics agree. OpenAI's embeddings guide notes that its embeddings are normalised to length 1, so cosine similarity can be computed as a plain dot product, and cosine similarity and Euclidean distance give identical rankings. Databases still make the choice explicit.

  • pgvector

    <=> is cosine distance, which is one minus cosine similarity. <#> returns the negative inner product, because Postgres only supports ascending index scans on operators. <-> is Euclidean (L2) distance.

  • Azure Cosmos DB

    Cosine is the default distance function, ranging from -1 to +1, alongside dot product and Euclidean.

Where cosine similarity misleads

  • Scores are not probabilities

    A similarity of 0.82 does not mean 82% relevant. Different models produce different score ranges, so a threshold tuned on one model breaks on the next.

  • It can be arbitrary

    Harald Steck, Chaitanya Ekanadham and Nathan Kallus showed in Is Cosine-Similarity of Embeddings Really About Similarity? (arXiv, 2024-03-08) that for some learned embeddings, cosine values depend on regularisation choices and can be arbitrary.

  • Magnitude is thrown away

    Cosine ignores vector length. For models that encode information in length, dot product is the metric the model was trained for.

  • Metric mismatch

    An index built for one metric and queried with another returns quietly wrong neighbours. The model card states which metric to use.

What it is confused with

  • Cosine distance

    Distance is one minus similarity, so lower is closer. Mixing the two flips a threshold's meaning.

  • Relevance

    Cosine measures closeness in embedding space, not whether a passage answers the question. A reranker judges relevance directly.

How Motiontic's engineers apply it

Motiontic's engineers use the metric each embedding model was trained with, and avoid fixed similarity thresholds unless they are calibrated on labelled data for that exact model. When thresholds are needed, for example in a semantic cache, they are re-tuned on every model change.

Cosine similarity, dot product and Euclidean distance compared

Cosine similarity, dot product and Euclidean distance compared
Cosine similarityDot productEuclidean distance
MeasuresAngle between vectorsAngle and length togetherStraight-line distance
Range-1 to 1Unbounded0 to unbounded
Closer meansHigherHigherLower
Uses vector lengthNoYesYes
On normalised vectorsSame ranking as the othersEquals cosineSame ranking as cosine
pgvector operator<=> (distance)<#> (negative)<->

Common questions

What is the difference between cosine similarity and dot product?

Dot product mixes direction and length; cosine similarity uses direction only. For vectors normalised to length 1, such as OpenAI embeddings, the two give the same result.

What is a good cosine similarity threshold?

There is no universal value. Score ranges differ by model, so any threshold has to be calibrated on labelled examples for the specific model in use.

What is the difference between cosine similarity and cosine distance?

Cosine distance is one minus cosine similarity. Similarity is higher for closer vectors, distance is lower. pgvector's <=> operator returns distance.

Should embeddings use cosine or Euclidean distance?

Use the metric the model card recommends. For normalised embeddings the rankings are identical, so the choice mainly affects how scores read, not which results come back.

Running agents that have to hold up?

Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.