Competency Questions as Executable Plans: a Controlled RAG Architecture for Cultural Heritage Storytelling

A research team led by Naga Sowjanya Barla and Jacopo de Berardinis has developed a novel neuro-symbolic architecture that addresses the critical challenge of factual accuracy in AI-generated cultural heritage narratives. Their paper, "Competency Questions as Executable Plans," introduces a controlled RAG (Retrieval-Augmented Generation) system that repurposes competency questions—traditionally used as design-time validation artifacts—into run-time executable narrative plans. This creates a transparent "plan-retrieve-generate" workflow that bridges the gap between high-level user personas and atomic knowledge retrieval, ensuring generation remains evidence-closed and fully auditable.

The researchers validated their architecture using the Live Aid Knowledge Graph, a multimodal dataset aligning 1985 concert data with the Music Meta Ontology and linking to external multimedia assets. They conducted a systematic comparative evaluation of three distinct RAG strategies: purely symbolic KG-RAG, text-enriched Hybrid-RAG, and structure-aware Graph-RAG. Their experiments revealed quantifiable trade-offs between the factual precision of symbolic retrieval, the contextual richness of hybrid methods, and the narrative coherence of graph-based traversal.

This work, accepted at the 23rd European Semantic Web Conference (ESWC 2026), offers actionable insights for designing personalized and controllable storytelling systems. By grounding generation in Knowledge Graphs and making the planning process explicit, the architecture provides a solution to the hallucination problem that plagues standard LLM approaches in domains where veracity is paramount, such as cultural heritage preservation.