Traditional agricultural advisory services face significant limitations in reaching smallholder farmers with timely, accurate information. Advancements in Large Language Models (LLMs) show potential for empowering agricultural extension systems, yet their direct application may pose risks due to lack of context-specific information. As farmers require informed decision-making support on diverse issues ranging from weather and agronomy to pest management and market dynamics, AI systems also need to be grounded in reliable, verified, and dynamic agricultural knowledge system.
The GAIA project aims to enhance the efficacy, reliability, and contextual relevance of AI-generated agricultural advisories for small-scale producers in the global South. Led by IFPRI with partners CABI, SCiO, University of Florida, and Digital Green, this project addressed critical gaps in agricultural extension services through AI applications.
During Phase I (2023-2024), GAIA generated key insights into AI-powered agricultural chatbot design and development. Through curated agricultural knowledge, pilot implementations, and research on data governance and gender bias assessment, the project demonstrated AI-driven advisory tools’ potential while identifying improvement areas. Testing a retrieval-augmented generation (RAG) framework with Digital Green’s Farmer.Chat in Kenya and India, the project examined how CGIAR’s open-access research and CABI’s proprietary materials could enhance AI-generated advisory accuracy and relevance. These efforts produced valuable learnings on user expectations, data strategy, performance evaluation, technical integration, and mixed-content governance, informing future agricultural AI applications.
Building on the foundational work of Phase I, GAIA Phase II (2025-2027) aims to further enhance the effectiveness and reach of generative AI-powered agricultural advisory services for small-scale producers through three key objectives. First, it will expand content aggregation beyond CGIAR sources while implementing robust data governance frameworks and developing a GenAI ethics toolkit. Second, it will enable dynamic advisories by integrating real-time data sources, predictive analytics, and multimodal models including crop health images. Third, it will establish comprehensive evaluation and benchmarking protocols to assess LLM performance in agricultural extension services, focusing on accuracy, timeliness, gender-sensitivity, and contextualization for diverse user groups.
