TokenPad

RAG

Cosine similarity

Cosine similarity measures the angle between two embedding vectors, giving a score from -1 to 1 for how similar their meanings are.

It compares direction rather than magnitude, which is what you want: a long document and a short sentence about the same topic should score as similar.

In practice scores cluster in a narrow band, so an absolute threshold like 0.8 is meaningless until calibrated against your own data and embedding model.

In practice

The scores are only comparable within one embedding model — 0.82 from one model and 0.82 from another mean nothing to each other. Absolute thresholds set by intuition also travel badly between corpora; pick the cutoff by looking at retrieved results on real queries rather than by choosing a round number.

Common questions

What similarity score counts as a good match?

There is no universal threshold. Scores are only comparable within one embedding model and one corpus, so set the cutoff by inspecting retrieved results on real queries rather than by picking a round number.

Why are all my similarity scores high?

Many embedding models produce a narrow band on real text, so 0.7 to 0.9 covers everything from perfect to irrelevant. What matters is the ranking and the gap between results, not the absolute value.

Measure it

  • EmbeddingAn embedding is a vector of numbers representing the meaning of a piece of text, so that similar meanings sit near each other in that space.
  • Semantic searchSemantic search finds text by meaning rather than by keyword, by comparing embeddings instead of matching words.
  • Vector databaseA vector database stores embeddings and finds the nearest ones to a query vector, usually with an approximate nearest-neighbour index.

More in RAG