OA-RAG: Ontology-based Agentic RAG for Faithful LLM Reasoning

OA-RAG integrates formal ontology reasoning with LLM generation to reduce hallucinations in domain-specific tasks. The agentic framework maps retrieval results to logical structures and outputs specific ontology observations alongside a four-tier grounding assessment: entailed, supported, related, or unrelated. This method improves faithfulness across six medical question-answering benchmarks.