Bibliographic Metadata Extraction Using Deterministic Retrieval and Specialized AI Agents

Abstract

Automated bibliographic metadata processing is a key component of intelligent library information systems. However, the use of large language models (LLMs) for bibliographic record generation is associated with factual errors, bibliographic hallucinations, and structurally incorrect metadata. This study proposes a retrieval-constrained workflow in which generative AI operates only on previously extracted and validated information from authoritative bibliographic sources. The workflow integrates deterministic metadata extraction, hierarchical source selection, normalization, multi-stage validation, and specialized AI agents for metadata structuring, Universal Decimal Classification (UDC) assignment, and MARC-formatted record generation. The architecture was evaluated on 200 real ISBN-identified bibliographic records representing 12 languages. The experimental design distinguishes ISBN-only diagnostic conditions, an external Google Books bibliographic metadata baseline, a retrieval-only internal architectural ablation, and the full specialized AI Library workflow. Google Books returned exact ISBN-matched records for 48 of the 200 submitted identifiers (24.0% coverage), resulting in a macro-average manually verified field-level accuracy of 14.8% over the complete dataset. In the controlled GPT-4o comparison, the specialized AI Library workflow increased the macro-average manually verified accuracy of the principal bibliographic fields from 65.0% to 83.2%. It also increased the UDC Recognition Rate from 85.0% to 99.5% and COMARC/B record-generation coverage from 87.5% to 97.5%, while mean Exact+Partial correspondence with the COBISS reference records increased from 62.1% to 73.6%. Overall, the results demonstrate consistent improvements in bibliographic metadata accuracy, UDC assignment coverage, and COMARC/B record-level correspondence.