One-sentence definition
Numerical representation that places elements in a space where certain relationships can be measured by proximity.
How to interpret this concept
An embedding represents an object via a learned numerical vector. In text search, the usual goal is for related queries and passages to be close according to a similarity measure. It is not a reversible translation of a phrase nor a certificate that two statements say exactly the same thing.
To compare vectors, the model and procedure used to generate them must be respected. Mixing embeddings from different models, even if they have the same dimension, does not by itself provide a compatible space. When changing models, it is usually necessary to re-embed the corpus and check retrieval quality.
Practical example: a query about how to cancel a purchase may retrieve a page titled order return even if it does not repeat the same words. However, an old policy or one from another country may also seem close. Combining similarity with date, language, and permission filters helps avoid responses that appear relevant but are inapplicable.
