Guide / cosine similarity in vector search

Cosine Similarity Explained for Vector Search

Cosine similarity measures direction between vectors. It is a useful ranking signal, not a correctness guarantee.

What cosine similarity measures

Cosine similarity compares the orientation of two vectors. If two normalized vectors point in a similar direction, their cosine similarity is high.

The common formula is cosine(a, b) = (a dot b) / (||a|| ||b||). Many embedding systems normalize vectors so cosine similarity and dot product become closely related.

Why direction can approximate semantic relatedness

Embedding models place texts with related statistical contexts near similar directions. That is why a query about login failure can retrieve a document about authentication errors.

The metric does not inspect truth, recency, permission, or operational safety. It only compares vectors produced by a model.

Cosine, dot product, and L2 distance

Cosine focuses on angle. Dot product can also include magnitude. L2 distance measures Euclidean distance. Which metric is appropriate depends on how the embedding model was trained and whether vectors are normalized.

Complexity Table

Metric	Measures	Good for	Risk
Cosine similarity	Vector orientation	Normalized embeddings	Ignores magnitude and correctness
Dot product	Alignment plus magnitude	Models trained for dot product	Magnitude can dominate
L2 distance	Raw coordinate distance	Geometric nearest neighbors	Scale sensitivity

When to Use This

Use cosine similarity when the embedding model expects angular comparison or normalized vectors.
Use it as candidate retrieval, then validate final quality with reranking and judged queries.

When Not to Use This

Do not treat high cosine similarity as proof that an answer is true.
Do not compare scores across different embedding models without calibration.

Production Failure Modes

High cosine similarity can retrieve text that is related but wrong for the user task. This is common with ambiguous terms, stale docs, and negated queries.

Thresholds can drift after an embedding model change. A score cutoff that worked on one model may fail after re-embedding.

Animated SVG Diagram

Two vectors point from an origin with a small angle between them, representing high cosine similarity.

Next Topics

vector search limitations semantic search vs vector search HNSW graph search vector database memory calculator production RAG retrieval architecture

FAQ

What does cosine similarity mean in vector search?

It measures how closely two embedding vectors point in the same direction, usually as a proxy for semantic relatedness.

Is cosine similarity better than dot product?

Neither is universally better. Use the metric recommended for the embedding model and verify retrieval quality on your data.

Does high cosine similarity mean the answer is correct?

No. It means the vectors are close by the chosen metric. Correctness needs grounding, filters, freshness checks, and evaluation.