LinkedCulture

Abstract

Cultural heritage discovery systems were designed for a retrieval environment that no longer fully describes the people who use them. The vocabulary of the cataloger and the vocabulary of the searcher are not the same, and no refinement of keyword search has been able to close that gap. This paper argues that the answer lies in a layered discovery architecture that combines keyword retrieval, semantic similarity, and AI-assisted query mediation, while preserving the authoritative institutional record as the only thing the end user ever sees. The central architectural principle is that AI belongs at the representation layer, not the interpretation layer. The system helps users find records. It does not generate, rewrite, or interpret them. LinkedCulture, an open-source prototype built across eight cultural heritage institutions and more than two hundred thousand records, demonstrates this architecture in operation and documents three observations from its deployment: that layered retrieval surfaces records keyword search alone misses, that neither retrieval mode dominates unconditionally across query types, and that a shared representational space appears to mediate vocabulary inconsistency across institutional boundaries in ways that single-institution search cannot replicate. The implications for how cultural heritage institutions approach discovery infrastructure, retrieval evaluation, and the appropriate role of AI in their systems are discussed.