In this article we describe a method for refining an initial set of search results, obtained through a meta-search engine, based on relevance feedback. The idea is to interactively obtain from the user a subset of relevant documents in an ongoing query, thereby providing a sample of the related vocabulary. Terms acquired in this way are combined with the terms initially in the query, in order to improve retrieval precision. In our method the user is also asked to select a subset of irrelevant documents, so that terms may be combined negatively in the query. A model of compatible architectures, in which the method can be implemented, is presented. An instance of such model, the system Web Query Reformulator (WQR), is described, with some of its performance results.