Re-ranking is a second scoring pass that reorders a small candidate set retrieved by faster methods, using a more expensive and more accurate model.
Retrieval optimises for recall across the whole corpus and must be cheap. Re-ranking optimises for precision across perhaps fifty candidates and can afford to be expensive — commonly a cross-encoder that reads query and document together instead of comparing precomputed vectors.
This two-stage shape is the standard architecture of modern search. It is also where business logic belongs: availability, margin, popularity and merchandising rules applied at re-rank time affect what is seen without distorting what was retrieved.