'When considering the development of musical digitization, new challenges emerge within the field of Music Information Retrieval, where our focus of research is querying on vast music collections. For that purpose we introduce and evaluate the Music On Demand framework where songs are queried as a continuous stream. When querying songs a listener is able to influence the songs ahead in the stream dynamically by performing the following actions: play similar songs, play random songs, skip songs, restrict collection and specify collection. In order to do so, a generic music data model and associated query functionalities are defined. Applying bitmap indices to index metadata as well as musical similarity derived from the musical content, we enable support for efficient retrieval within vast music collections by the use of bit-wise operations. The retrieval process concerns a combination of both the metadata and the similarity of songs. In this context we examine the use of the Word-Aligned Hybrid compression scheme and the Attribute Value Decomposition technique for representing content based similarity. Experimental test results show that our framework implementation ensures efficient access to music within vast music collections, at the cost of only a small additional space consumption when compared to the stored music files. '