Relevance feedback is particularly powerful when used in conjunction with ProbabilisticSearching. Initially, query terms are given a "weight" in inverse proportion to their frequency, on the grounds that the rarer terms will be better predictors of relevance. Whe the user provides relevance feedback, i.e. states whether a document is useful or not, new terms are extracted from the relevant document, and they and the original terms are re-weighted taking relevance into account, so as to form a new query which
better represents the user's information need. That's the theory anyway, and in practice the method has proved effective in competitive situations, e.g. TREC.