Retrieval-grounded multilingual LLM assistance for island smallholder farmers.

Abstract

Background: Smallholder farmers on remote, depopulating islands have limited access to advisory services, and their locally specific agronomic knowledge, often expressed in regional dialect, is poorly represented in the corpora on which large language models are trained. A general-purpose chatbot therefore answers local questions fluently but unreliably. A conversational artificial intelligence assistant, Falco eleonorae, was developed for the farmers and cooperatives of Kythera and Antikythera, Greece.

Methods: The assistant is a thin Backend-for-Frontend proxy in front of a managed, geospatially aware agronomic agent, hosting no chat model of its own. Answer generation and tool selection are delegated to upstream models; a vision-capable model turns an uploaded field photograph into a short description, so only text reaches the agent; voice is transcribed by a managed speech-to-text service in a European Union region. Grounding uses tool-augmented retrieval rather than a self-hosted vector database: a Model Context Protocol tool queries a curated, read-only, bilingual data interface, and every request carries a geospatial Well-Known Text envelope anchoring the agent to the islands.

Results: It is deployed as a Greek-primary progressive web application supporting text, voice, and image input with streamed responses, suited to low digital literacy and poor connectivity. Locally scoped questions are answered from the pilot's editorially controlled records rather than the model's parametric memory, and content edits take effect on the next query without re-indexing. Safeguards include European Union data residency for speech, server-side editorial gating, rate limiting, output sanitisation, and erasure of chat content on account deletion.

Conclusions: For a small, resource-constrained rural deployment, a managed and grounded multilingual assistant is more attainable and trustworthy than a bespoke self-hosted model. No formal user evaluation has yet been performed, and a structured field evaluation is the principal future work.