The context of the REMUS project is a just-in-time document recommendation system, which spontaneously retrieves and displays relevant information to users engaged in a conversation such as a business meeting. In previous work in the IM2 NCCR, we have designed a keyword extraction method constrained by topic models, which builds multiple topically-independent implicit queries. The goal of REMUS is to design, and then evaluate, advanced algorithms for merging and re-ranking the results of multiple queries. These algorithms will reward the diversity of results in order to maximize the coverage of all the queries and their possible topics with the smallest possible number of documents. Moreover, REMUS also aims to re-rank the results of spoken queries explicitly directed to the system. The increase in the relevance of the document results will be measured by human assessments obtained by crowdsourcing using a protocol and metric that we have previously defined. The algorithms will improve the accuracy of the approach and advance the state of the art.