Agricultural large language model for standardized production of distinctive agricultural products

Abstract

To address the diverse nature of specialty agricultural product standardization, its complex and cumbersome development process, and lengthy drafting cycles, while simultaneously tackling challenges such as outdated standardization documents and hallucinations caused by general large language models’ delayed access to agricultural domain information. This study constructs a multi-stage cascaded large language model based on a hybrid retrieval-augmented mechanism. The model comprises three core modules: (1) A multi-source retrieval augmentation module that achieves comprehensive external knowledge acquisition through vector retrieval, keyword retrieval, and knowledge graph retrieval branches; (2) A knowledge fusion module that filters redundant information using inverse ranking fusion and graph structure pruning methods to achieve precise injection of high-quality knowledge; (3) A domain adaptation module that enhances the model’s understanding of agricultural terminology through vertical domain fine-tuning. Experimental results show that in the standardization document summarization task, the model achieves chrF, BERTscore, and Gscore metrics of 34.85, 74.88, and 39.85, respectively, representing improvements of 59.52%, 35.28%, and 72.84% over the BART baseline model, and 58.54%, 24.25%, and 59.54% over the T5 model. This study enriches the theoretical foundation of large language models in agriculture and provides intelligent technical support for specialty agricultural product standardization development.