https://drive.google.com/file/d/1lZAFsK7zzak4SSasSAhEYuCMLQawAnBA/view?usp=share_link 

Healthcare is one of the most demanding real-world environments for knowledge-driven AI: data sources include ontologies, biomedical literature, clinical narratives, and regulated patient records that cannot be freely combined or centralized. While Retrieval-Augmented Generation (RAG) has improved LLM grounding, traditional text-centric pipelines fall short when explainability, temporal reasoning, and governance are first-class requirements.

This session presents a GraphRAG architecture for regulated domains, with healthcare as a case study. The approach uses materialized medical knowledge graphs to unify ontologies, publications, and guidelines, while patient-level EHR data remains virtually integrated through controlled access rather than ingestion. This separation preserves compliance while enabling contextual reasoning.

The architecture introduces LLM-based agents as rational planners that decompose questions into graph queries, invoke rules and constraints, and orchestrate reasoning across the knowledge graph. Graph Neural Networks (GNNs) support relevance ranking, similarity scoring, and subgraph selection, enabling scalable inference beyond symbolic traversal.

The workshop connects core conference themes: knowledge graphs and ontologies as AI infrastructure; rules and reasoning to constrain and explain LLM behavior; agents as graph-centric orchestrators; and enterprise concerns, including governance, observability, and data virtualization.