The idea
An embedding maps input to a numeric vector using a particular model. Related inputs may lie near each other under a chosen similarity measure. This is a learned representation, not a verified statement of truth. Vectors from different models or versions are not automatically comparable.
Worked example
A query about sending back a broken item may retrieve a returns-policy paragraph even without the same keywords. A nearby shipping paragraph might still be irrelevant. A high similarity score means the representation finds them related, not that either answers the question.
Try it
Write four fictional support paragraphs and three differently worded queries. Rank paragraphs manually by relevance. Identify a result that shares keywords but fails the actual question, and explain why an embedding score would still need evaluation.