ENHANCING AGRICULTURAL DECISION-MAKING WITH COMBINED LANGUAGE MODELS

Abstract

Agricultural advisory systems play a crucial role in supporting farmers by providing timely and accurate solutions to crop-related problems. However, building reliable question-answering systems for agriculture is challenging due to the diverse, domain-specific, and often noisy nature of farmer queries. In this project, a retrieval-based agricultural questionanswering framework is proposed to deliver accurate and trustworthy responses by leveraging historical expert-curated data. The system utilizes a pretrained sentence embedding model to convert farmer queries into dense vector representations, followed by similarity-based retrieval using FAISS to identify the most relevant past questions and their corresponding expert answers. Unlike generative models, which often produce vague or hallucinated responses, the proposed approach ensures factual correctness by reusing validated advisory content. Experiments conducted on a large farmer query dataset demonstrate that the system achieves a training accuracy of 100% and a test accuracy of 95.9%, with qualitative evaluation confirming high relevance and agricultural validity of the retrieved responses. The results indicate that retrievalbased question answering is a practical, efficient, and reliable solution for real-world agricultural advisory applications.