Business intelligence is crucial for synergy and competitiveness in the world of business; the challenge is to capitalize upon the experiences, relationships and knowledge of all members over time. In this paper, we address business intelligence and more generally collective intelligence through a social approach based on viewpoints. The collective knowledge is stored within a bi-partite graph populated by agents, documents and topics on one side and by viewpoints on the other side. Viewpoints consist in labelled triples (agent, document, topic). We define a semantic distance on this structure. We then engage a selectionist process: information retrieval is based on existing viewpoints while feedbacks yield new viewpoints. The implemented framework and algorithms are tested through a simulation.