NeuroRAG: A Multi-Agent Model for Neurobiological Question Answering

In biomedicine, answers must be complete, current and verifiable, yet a general-purpose large language model (LLM) can give confident but wrong or outdated answers with no sources to check. We present NeuroRAG, a multi-agent retrieval-augmented system that answers neurobiology questions using retrieved scientific sources. Its knowledge base combines a hierarchical index of neurobiology textbooks with live search of PubMed, arXiv, bioRxiv/medRxiv and NCBI Gene/Protein sources, and web search as a fallback. A router selects the sources, query-rewriting agents reformulate the question, passages are fused, reranked and compressed, three LLMs draft answers, which are merged, and, as the last step, graders check grounding in the retrieved text and relevance to the question. On benchmarks, NeuroRAG had the highest accuracy on seven of nine MMLU subsets; on PubMed Summary QA, its similarity to reference answers was comparable to Llama 3.3 70B and lower than GPT-4o and GPT-4.1. Against two published RAG pipelines, it beat FlashRAG Naive RAG on six of seven metrics and was indistinguishable from MedRAG. Because such benchmarks are limited to text and age quickly, we ran a blind evaluation of the system by 14 evaluators: NeuroRAG was chosen in 64.4%, GPT-5 in 14.1%, Claude 4.5 Sonnet in 11.0%, and none in 10.4%. This shows expert preference for NeuroRAG over the other models, though not a direct measure of factual correctness.