Our claim is that semantic metadata are required to allow a real reusing and assembling of learning objects. Our system is based on three models used to describe the domain, learners, and learning objects. The learning object model is inspired from knowledge representation proposals. A learning object can be reused directly or can be combined with other learning objects using composition operators with well defined semantic. Using these models we are able to define powerful search tools and an adaptive environment taking as input the learner model to construct the learning object to deliver. We are currently implementing this system using Sesame, an RDF repository.