Off-line LLM-based decision tool to help advise Farmers in sub-Saharan Africa for Enhanced Decision Making in Goat Care

The objective of this work is to assist goat farmers based in Sub-Saharan Africa in determining whether veterinary consultation is necessary when their animals are unwell, thereby helping to improve the farmers ability to care for their animals and also avoid unnecessary costs. The study presents a prototype offline mobile application that guides users through a process using a decision tree and user-provided responses. The work has investigated the potential of large language models (LLMs) for generating decision trees and diagnostic tools in data-scarce domains such as veterinary medicine. Retrieval-Augmented Generation (RAG) was employed to enhance result quality and accuracy. Decision trees were generated with the Gemma 32B-parameter model; due to complexity and size limitations, large trees were constructed incrementally with carefully designed prompts. Diagnostic generation used the DeepSeek 14B-parameter model, which consistently produced the most precise and reliable outputs. Key limitations included hardware performance—directly affecting computation time as a function of model size and machine specifications—and the models’ tendency toward overly cautious answers. To partially mitigate these constraints, techniques such as trigger words, verification prompts, and scoring methods were applied under specific conditions. Nevertheless, expert validation remains necessary, as these verification methods cannot be considered 100% reliable.