The paradigm of Content-based retrieval (CBR) with query-by-example (QBE) has been prevailing in multimedia applications and systems. While the CRB/QBE approach is easy to use from the end-user's perspective, it is not always possible for a user to prepare/have a suitable "example" in hand; besides, the lack of semantics inherent to this approach may cause strange things to happen (e.g., funny results). In this paper, we argue that multimedia retrieval should be knowledge-based and semantics-driven, preferably with personalization support. Through a case study on a distributed multimedia recipe database (DMRD) system, we show that User-community Profiling + Domain Knowledge + CBR (including keywords/QBE) is an effective way to go forward. Experiments on a prototype of DMRD system demonstrate the feasibility and effectiveness of such a hybrid approach we are advocating.