Item

Abstract

Rapid growth and increasing fragmentation of Earth-system science literature pose a major barrier to timely, evidence-based assessment of global environmental change and planetary boundary conditions. Here we present a domain-adapted retrieval-augmented generation framework for Earth-system research, in which domain adaptation is achieved through curated scholarly corpora and a structured evaluation design, supporting transparent and reproducible deployment. The system continuously curates literature from major scholarly databases and integrates it with locally hosted, open-weight language models to enable source-level traceability, long-term accessibility, and independence from external commercial services. When queried, it retrieves relevant scientific sources and generates responses grounded directly in those documents, reducing screening effort and addressing reliability concerns associated with general-purpose language models. Quantitative evaluation shows high contextual faithfulness (0.96), strong answer relevance (0.88), and robust correctness (0.81). Our findings indicate that domain-adapted retrieval-augmented generation can strengthen the scientific basis for Earth-system assessment by enabling transparent, continuously updated, source-attributed synthesis across rapidly expanding literature.