Agentivité épistémique et recherche assistée par IA: un système de recommandation transparent pour la recherche informationnelle exploratoire

Information retrieval tools in academic contexts filter and rank what researchers read and cite, thereby influencing scientific innovation and the citation dynamics that shape research. Current “AI research assistants” rely on a natural language querying strategy interpreted by a large language model ; this paradigm, which appears to attract the majority of information retrieval resources today, raises questions about the epistemological impact of interpretations performed by tools whose black-box effect is well-documented. This paper presents a recommender system appended to the specialized scientific search engine Isidore, tool designed as a concrete alternative to natural language querying. First, we will present the functionalities of the recommender system developed in this project: constructing a query through iterative interactions involving terms and definitions using the IEML language; comparing result lists displayed side-by-side from the user-constructed query and the same query “augmented” by an LLM. Second, we will present the evaluation of the semantic parsing into IEML performed by an LLM on a corpus of 360 terms, as well as the results of a small sample user study.