A Tool-Augmented AI Agent for Grounded Access to European Cultural Heritage

Large Language Models (LLMs) have demonstrated impressive conversational abilities, yet they remain prone to hallucination, the generation of plausible but factually incorrect information. This limitation is particularly critical in the cultural heritage domain, where the accuracy and reliability of historical and artistic data are fundamental requirements. In this paper we present a tool-augmented AI agent designed to answer natural-language queries about European artists and artworks by grounding every response on verified metadata retrieved from the Europeana digital platform. The system integrates the CrewAI agent framework with a GPT-family language model and the Europeana REST API, constraining the model to a single external tool call per conversational turn to ensure deterministic and verifiable behaviour. A lightweight short-term memory mechanism enables coherent multi-turn interactions without context contamination. Empirical evaluation on a systematic benchmark of 150 queries spanning 50 European artists demonstrates perfect tool discipline (100.0%), near-universal source adherence (98.7%), and complete temporal coverage (100.0%), compared with 87.3% temporal coverage for an isolated LLM baseline that cannot cite verifiable sources. The Streamlit-based interface exposes technical provenance details to the user, fostering transparency and trust. Our results confirm the viability of tool-augmented generation as a mediator between structured cultural archives and lay audiences, contributing to the democratisation of digital heritage in the Digital Humanities field.