Integrating LLMs into MARC 21 Cataloguing: A Human-in-the-Loop Approach within Koha ILS

This study examines the integration of large language models (LLMs) into the Koha Integrated Library System (ILS) to automate selected cataloguing functions while ensuring compliance with MARC 21 standards. A prototype system was developed to support translation of titles, summary generation, Library of Congress Subject Headings (LCSH) assignment, and Dewey Decimal Classification (DDC) selection. The system incorporated structured prompt engineering, deterministic inference settings, and a human-in-the-loop design to maintain metadata accuracy. Ten MARC 21 book records from Springer Nature collections were processed, and LLM-generated outputs were evaluated against professionally curated records using cosine similarity, BERTScore, and DDC prefix match. Results showed high alignment with reference records, with mean scores of 77.96% for summaries, 73.92% for keywords, and 89.65% for DDC prefixes. Hindi and Bengali translations achieved average BERTScores of 80.25% and 79.41% respectively, indicating strong multilingual fidelity. The findings suggest that LLM-assisted workflows can enhance metadata completeness, consistency, and discoverability, particularly for underutilized MARC 21 fields, while reducing manual effort. Limitations include the small sample size, domain-specific data sources, and reliance on external API services. The study contributes evidence on the operational viability of AI-assisted cataloguing and offers a framework for balancing automation with professional oversight in library metadata creation.