Information brokerage systems support distributed information gathering for data collections stored within the World Wide Web or for other on-line data repositories. These systems present a unified face of the underlying heterogeneity, both in terms of the access/search protocols and in terms of the data schemata. Typically, the user may input some values (e.g., interpreted as strings within a fulltext search), and may refine/relax them once she gets too many/little results for her query. Adaptiveness with respect to a refinement of attributes can improve the precision of the searches dramatically. This paper discusses adaptive refinement of attribute-patterns together with its main implications in the context of the Constraint-Based Knowledge Broker system. It presents an implementation for distributed information gathering, and motivates design choices through concrete examples.