Auto-RAG: Hybrid Retrieval-Augmented Generation with conditional routing for nutrition knowledge access

Large Language Models (LLMs) have demonstrated strong capabilities in natural language generation; however, they remain prone to hallucinations and often lack verifiable grounding when addressing knowledge-intensive queries. Retrieval-Augmented Generation (RAG) alleviates this limitation by incorporating external knowledge sources, yet most existing approaches apply retrieval indiscriminately, which can introduce irrelevant context and increase computational overhead. In this paper, we propose Auto-RAG, an adaptive retrieval-augmented generation framework with conditional routing that dynamically determines whether external knowledge retrieval is necessary for a given query. The framework employs a lightweight query analysis module to assess query complexity, intent, and lexical specificity, enabling intelligent routing between a direct LLM generation pathway and a retrieval-augmented reasoning pipeline. To enhance retrieval effectiveness, the proposed approach combines semantic embedding-based search with keyword-based BM25 through a unified fusion and re-ranking mechanism, ensuring high-quality and contextually relevant evidence selection. In addition, we introduce a trust-aware evaluation framework that jointly measures answer quality, faithfulness to retrieved evidence, hallucination rate, and retrieval efficiency. Experimental results on a nutrition knowledge corpus derived from Human Nutrition demonstrate that the proposed framework achieves a strong balance between fluency and factual reliability. Compared to conventional always-on retrieval strategies, Auto-RAG reduces retrieval usage by approximately 30% while maintaining competitive answer quality and evidence-grounded performance. The results also show that the effectiveness of evidence filtering depends on its configuration, with overly aggressive filtering potentially removing useful supporting evidence. While LLM-only responses may exhibit higher linguistic fluency, they lack grounding guarantees, underscoring the importance of adaptive retrieval in safety-critical domains. Overall, this work presents an efficient and reliable approach for domain-specific knowledge access and highlights conditional retrieval as a scalable paradigm for improving trustworthiness in large-scale language systems.