Abstract: To address the issues of insufficient professional knowledge constraints, limited long-chain reasoning capabilities, and possible factual hallucinations in existing general large language models for agricultural pest and disease diagnosis, this study proposed an Agent-based dual-channel intelligent diagnostic system integrating knowledge graphs and large language models. At the data level, a Triple-Aware Efficient Denoising (TAED) mechanism was designed to eliminate low-quality redundant data and construct a corpus for domain-specific fine-tuning. At the model level, a Phased-Adaptive Chain-of-Thought (PA-CoT) strategy was proposed. Through instruction fine-tuning, the model learned diagnostic logic and reduced the risk of hallucinations caused by logical jumps. At the architecture level, an intent recognition-based Agent dual-channel retrieval mechanism was constructed to achieve the complementary enhancement of unstructured semantic information and structured knowledge. Experimental results demonstrated that the proposed system achieved an accuracy of 82.4% and a hallucination rate of 10.8% in complex agricultural diagnostic tasks. It outperformed traditional Retrieval-Augmented Generation (RAG) methods in terms of reasoning completeness and knowledge accuracy, providing a solution for the practical application of agricultural intelligent expert systems.

Key words: agricultural pest and disease diagnosis, large language models, knowledge graph, dual-channel retrieval