This paper studies a reference network architecture for video information retrieval, based on hierarchical storage of the video segments and asynchronous transfer mode (ATM) transport capability on the long distance network. The storage associated to local switching nodes can be exploited in two complementary ways, either for caching the most requested video programs, and for decoupling the transmission speed on the long distance network from the natural playout speed of the video segments. We show that both approaches lead to a substantial increase of the load carried on the long distance network with a given service requirement. Moreover, we show that a greater efficiency is obtained when a suitable integration of the two approaches is considered.