CAR-RAG: Category-Aware Hybrid Retrieval-Augmented Generation for Hallucination Mitigation

We introduce CAR-RAG (Category-Aware hybRid Retrieval-Augmented Generation), an approach to mitigate hallucinations in large language models (LLMs) for real-world deployments. Unlike single-modality pipelines, CAR-RAG conditions retrieval on semantic query categories and adaptively combines vector retrieval with lightweight causal query augmentation. This category-aware routing improves risk profiles by reducing contradicted claims, which are particularly harmful in safety-critical settings.

We evaluate CAR-RAG on an automotive Q&A dataset comprising over 700 community-provided questions and answers from Stack Exchange's Motor Vehicle Maintenance & Repair forum. The framework achieves approximately 90% factual accuracy and reduces confidently incorrect statements to 6.4%, outperforming dense retrievers and the base LLM in risk-sensitive settings.

These results highlight trade-offs between peak accuracy and robustness and position CAR-RAG as a practical, interpretable, deployment-ready solution for hallucination mitigation. It suits industrial contexts (automotive troubleshooting, service-center support, and technical diagnostics) where high-precision answers are essential. An open-source implementation is available on GitHub.