OA-RAG: Ontology-based Agentic RAG for Faithful LLM Reasoning
Summary
- Researchers propose OA-RAG, an ontology-based agentic retrieval-augmented generation framework designed to improve LLM reasoning faithfulness in sensitive domains like biomedicine.
- OA-RAG integrates formal logic and domain ontologies with LLM reasoning, providing explicit evidence observations and model-assessed grounding levels for generated answers.
- Evaluated across three large-scale ontologies and six real-world medical QA tasks, OA-RAG outperforms popular RAG methods and competitive domain-specific textbook RAG systems, especially on reasoning tasks.
Why it matters
By combining formal logical ontologies with agentic RAG, this approach provides verifiable evidence and grounding levels that mitigate hallucinations in critical domains like healthcare.