An LLM-based Multi-agent Framework for Publishing RDF Views into Enterprise Knowledge Graphs

Abstract

The publication of heterogeneous data into Enterprise Knowledge Graphs (EKGs) remains challenging due to schema matching, semantic inconsistencies, and the effort required to create mappings and target ontology. This paper presents a metadata-driven multi-agent framework based on Large Language Models (LLMs) for generating RDF views from data sources. The framework orchestrates agents for schema extraction, ontology bootstrapping-refining, semantic mapping generation, and RDF materialization tasks. The approach leverages metadata, represented through the Vocabulary of Semantic Views (VoSV) into a data design pattern (DDP SV), as in-context learning to support explainability, governance, and incremental publication. As a way to perform a preliminary evaluation of this framework, a case study using Brazilian public government purchases data sources demonstrates the feasibility of automating RDF view publication.