Evaluation of Extended Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) for Technical Assistance in Agriculture: Virtual Assistant to support the Agricultural Sector

Iparraguirre-Villanueva O.; Robles-Espiritu W.; Epifanía-Huerta A. · International Journal on Advanced Science, Engineering and Information Technology (2026) · DOI: 10.18517/ijaseit.16.2.21472

Resumen

Modern agriculture faces ever-increasing challenges that require access to accurate and timely information to optimize yields and manage problems. It is important to note that weather patterns no longer follow the same patterns, and pests have shown an ability to adapt, complicating the control process. This has an impact on those who work on the land, forcing them to make quick decisions to save their crops. While it is true that there is a lot of information available and scattered throughout scientific documents, this limits the possibility of integrating data to solve problems and adjust to real cases. This paper details the development of a virtual assistant application that uses Large Language Models (LLM) to provide intelligent and personalized assistance in the agricultural domain. The methodology involved the collection and curation of an extensive corpus of agricultural data and the selection and adaptation of LLMs, evaluating models such as GPT-3.5-Turbo, Llama 2 (7B) and Gemini 1.5 Pro. A Generation Augmented Retrieval (GAAR) architecture was implemented to anchor LLM responses in the curated knowledge base, significantly improving accuracy and reducing hallucinations. Preliminary results show that Model C (Gemini 1.5 Pro) with RAG achieved the best accuracy (93%) and lowest frequency of hallucinations (3%), outperforming previous expert systems and chatbots by providing contextual and reliable responses. The findings highlight the potential of this tool to democratize agricultural knowledge and improve decision making, with future plans to expand the corpus, integrate multimodal capabilities, and conduct field impact assessments. IJASEIT is licensed under a Creative Commons Attribution-Share Alike 4.0 International License.

Instituciones: Universidad Tecnológica del Perú; Universidad César Vallejo