Vector Similarity
Also called Similarity Measure
The math for how "close" two embeddings are — the thing semantic search actually ranks by.
Think of it like
Measuring how alike two colors are by how near they sit on the color wheel.
Example
A search ranks chunks by cosine similarity to the query embedding and returns the top few closest.
How it actually works
Once text is embedded, "relevance" becomes geometric distance. Cosine similarity (angle), dot product (angle and magnitude), and Euclidean distance are the usual measures, and the right one depends on whether your vectors are normalized. Retrieval quality can hinge on this choice, and mismatching the metric to the model is a classic silent bug.
For product teams
An invisible but consequential setting; the wrong metric quietly degrades every search.
For engineers
Distance/similarity function (cosine, dot, L2) over embeddings; must match the model’s training and normalization.
Related
- Cosine similarity — Cosine similarity is the common choice.
- Dot Product Search — Dot product is a faster variant.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome