MILHOUSE: LLM-basierte Sacherschließung in der Praxis

Ein Werkstattbericht der Universitätsbibliothek Magdeburg

Abstract

Academic libraries are experiencing a rapid increase in media holdings, making comprehensive subject indexing increasingly difficult to manage with existing staff resources.
This paper presents MILHOUSE, a tool developed at the University Library of Magdeburg that leverages large language models for the (semi-)automated assignment of subject classification notations.
The application enriches bibliographic records with suitable metadata, employs structured prompts with fixed output schemes, and integrates a plausibility check to reduce model-induced hallucinations. The quality of the results is ensured through a human-in-the-loop approach.
MILHOUSE was practically tested by subject librarians of the University Library of Magdeburg in a near-production setting (Q3 2025).
The paper highlights the advantages of this lightweight approach over traditional training pipelines and derives design principles for the rapid and responsible integration of artificial intelligence into library workflows.