Context-aware citation suggestion: A retrieval-centric approach for academic writing

Accurately recalling and citing relevant prior research remains a persistent challenge in academic writing, particularly in the science and engineering domains. This study presents a citation suggestion system based on Retrieval-Augmented Generation (RAG) to reduce the cognitive burden associated with manually locating and verifying references. We evaluated the RAG-based system on 120 research papers, primarily from battery science with pedagogical sciences as a comparison domain, comparing its performance to that of direct Large Language Model (LLM) prompting without a preloaded reference repository. The direct prompting approach exhibited severe hallucinations and failed to exceed 10% citation accuracy, rendering it unsuitable for academic use. In contrast, the RAG-based method achieved substantially higher accuracy—up to 61.0%—significantly surpassing the random baseline of ∼16%. These findings demonstrate that context-aware retrieval substantially enhances citation reliability, especially in disciplines with structured citation practices. The results underscore the value of retrieval-based approaches in Artificial Intelligence (AI)-assisted academic writing, offering a scalable solution to improve reference precision and reduce manual effort.