Admission assistants built on large language models answer applicant questions from a mix of regulations, programme descriptions and deadlines that change every intake. We describe a retrieval scheme that indexes admission documents together with the curriculum they belong to and re-ranks passages by the programme the applicant asks about. On a set of 1,240 real questions collected during two intakes the scheme raises answer accuracy from 71 to 86 per cent and cuts answers that cite a superseded regulation from 14 to 3 per cent. We report the failure cases that remain, most of which come from questions that span two programmes at once.