The various relevant aspects for the retrieval of images present familiar documentation issues as well as some entirely new requirements. The Semantic Index System (SIS) has already been successfidly employed in maintaining an evolving body of knowledge about the cultural and historical contexts of museum artifacts and multimedia documents (CLIO SYSTEM). Its distinct features of runtime meta-modelling and classification of references satisfy the multiple classification needs for images and their parts and allow to access simultaneously through detailed annotations. Consequently, it integrates access methods through generic and specific indices and the relevant context. In this paper we present the principles and architecture of such a system as it is currently being implemented.