Daejeon, July 27–31
This study proposes a Graph-RAG-based AI Smart Docent web prototype that integrates a semantic knowledge graph with generative AI to overcome both LLM hallucination and the structural constraints of physical signage. Current cultural heritage information services face two structural limitations. First, traditional physical interpretive media remain confined to unilateral information delivery, providing uniform content despite the diversity of visitors’ ages, prior knowledge, and languages. Second, general-purpose LLMs introduced as an alternative tend to generate factually inaccurate information due to their lack of domain-specific knowledge, thereby undermining the epistemic reliability required in cultural heritage interpretation.
This study selects the Hanging Painting of Janggoksa Temple (Maitreya Buddha, 1673, National Treasure) as a single empirical case. The work depicts dozens of iconographic figures—buddhas, bodhisattvas, and arhats—on a single canvas, and its hwagi (畫記) explicitly records diverse donation categories (materials, pigments, ritual offerings, production stages) and a roster of participating monk-painters. This data composition, in which iconographic classification and person-relation information richly coexist, provides suitable conditions for validating the proposed integration of a thesaurus and a relation-expressive knowledge graph (Jeong 2012; Park 2012; Korea Heritage Service 2022). The methodology is first validated through this work and then progressively extended to Buddhist sculpture, architecture, and archaeological sites, establishing a foundation for a generalizable generative docent model across Korean cultural heritage.
This study formalizes humanities data—basic specifications, hwagi records, personal records, and thesauri—into RDF/OWL-based semantic data as a representational layer for scholarly knowledge sharing, and converts these data into a graph database for Graph-RAG inference. A generative AI Smart Docent model is then developed using verified data retrieved through natural language queries, and implemented as a web-based prototype.
This study builds on prior research that leverages digital technologies to expand museums beyond their physical boundaries, including Semantic Web-based meta-archives (Jung 2023), immersive digital exhibitions in ubiquitous environments (Yoo 2014), and personalized smart docent services (Gu / Shin 2016). While generative AI has recently enabled highly personalized information delivery without pre-defined scenarios, its hallucination problem remains a critical limitation in domains requiring historical accuracy (Natale et al. 2025). Therefore, recent studies (MuseRAG++, Cultural Heritage Assistant) emphasize grounding responses in expert-verified data to ensure responsible curation and reliability (Trichopoulos et al. 2023; Nazar / Acosta 2023; Wang et al. 2025; Hu 2025).
The core of this study lies in restructuring analog and unstructured source resources into machine-readable forms inferable by generative AI, thereby technically mitigating AI hallucination. (1) The signage guidelines issued by the Korea Heritage Service are formalized as rule-based constraints embedded into the system prompt, specifying a priori the narrative principles the AI must observe (Korea Heritage Service 2024). (2) Four categories of heterogeneous data—basic specifications, hwagi records, personal records, and thesauri—are formalized as RDF/OWL ontologies, leveraging SKOS for the iconographic thesaurus and CIDOC-CRM for semantic relations among persons, works, places, and events. The formalized data are then converted and loaded into a property graph in Neo4j for Graph-RAG inference, with iterative refinement: load results are validated via Cypher and the RDF/OWL source is revised on discrepancies. This establishes a dual structure: RDF/OWL functions as the representational standard for knowledge modeling and scholarly sharing, while the Neo4j property graph supports real-time inference and query response. (3) Multilingual designations and alternative names (imyeong, 異名) are collected and mapped from authoritative sources including the Digital Dictionary of Buddhism (Muller 2009), the Encyclopedia of Korean Culture, the Korean Heritage Portal, and the hwagi of the Janggoksa Hanging Painting of Maitreya Buddha (Korea Heritage Service 2022). Each value receives a source-identifying prefix to secure scholarly accuracy and provenance, technically supporting the epistemic reliability of AI-generated interpretation.
Figure 1. System Architecture of the Interactive AI Cultural Heritage Docent
The Graph-RAG of this study is a RAG variant that uses structured query results from the graph database as grounding sources, aligning LLM responses with verified facts (see Figure 1). The backend server classifies natural language queries by type and selectively executes pre-defined Cypher queries, reliably retrieving core humanistic elements such as participating artists, dates, iconography, and hwagi records. In the generation stage, the LangChain framework integrates retrieved graph data into the prompt, with Google Gemini as the generative model. The system prompt incorporates the Korea Heritage Service guidelines, and the system adjusts narrative style and difficulty according to user expertise (children, general audiences, domain experts) and language preference, forming a customized generation mechanism in which the same grounding data can be reconfigured for different contexts.
Implemented with a Next.js front-end and a FastAPI back-end, the prototype offers a button-based interface for question selection and structured, data-grounded interpretation. The study further proposes an AI Smart Docent function in which core iconographic elements serve as SVG-based clickable images; selecting an element triggers real-time retrieval of associated graph data and generates a customized interpretation. This multimodal approach—integrating visual, textual, and structured data—lays a foundation for future extension to AR and 3D-based exhibition environments responsive to visitor interaction.
This study recontextualizes professional docent interpretation as a visitor-driven interactive experience. By formalizing dispersed humanities knowledge into a dual structure of RDF/OWL-based representation and graph-database-based inference with integrated real-time access, the system simultaneously supports scholarly knowledge sharing and AI-driven inference. By using verified semantic data as a grounding source to mitigate AI hallucination, the study presents the technical feasibility of responsible curation in digital environments. Ultimately, the model contributes a generalizable framework extensible to diverse cultural heritage domains, advancing the practical use of digital heritage and scholarly methodologies. The system's effectiveness, however, requires further validation through user evaluation and quantitative testing; the constructed RDF/OWL ontology and implementation code will be released with a forthcoming journal article to support reproducibility and collaborative extension within the digital cultural heritage research community.