Cosine similarity

Cosine similarity computes the cosine of the angle between two vectors; the result lies between -1 and 1, and closeness to 1 means greater directional similarity. What counts is the direction the vectors point in, not their length.

That is the desired behaviour in text search. A long document passage and a short question produce vectors of different magnitude; Euclidean distance penalises that difference, while cosine similarity looks only at semantic direction and ignores length.

Cosine scores are not an absolute quality measure. A score of 0.82 may be excellent in one archive and mediocre in another; the value depends on the embedding model and the nature of the content. That is why relative ranking is used instead of a fixed threshold.

The practical consequence: the system retrieves the few highest-scoring passages, not those above an absolute cutoff. If nothing relevant exists in the archive, something is still returned — which is why the model's ability to say "there is no answer in this context" must be defined as separate behaviour.

Vector databases usually support several distance metrics: cosine, inner product, and Euclidean. On normalised vectors, cosine and inner product produce the same ranking, which is why many systems normalise vectors and use inner product.

The choice must match the embedding model. Whichever metric the model was trained for is the one to use; a mismatched metric causes a silent loss of precision that is hard to notice because it raises no error.

A useful way to evaluate cosine-similarity is to follow one real question from the source document to the final answer. Record which file was selected, what context reached the model, and what a reviewer would need to verify. This turns a definition into an operational check and makes the result comparable across teams.

The same check should include ownership and change management. Decide who updates the relevant documents, how an outdated result is reported, and which access boundary applies. KobiGPT can provide the assistant and the cited document context, but the organisation still owns the source material, permissions, and the decision made from the answer.