A Retrieval-Grounded Multilingual Advisory System for Postharvest Decision Support in Low-Resource Settings: Design and Evaluation

Retrieval-Grounded Multilingual Postharvest Advisory for Low-Resource Settings 1 Introduction Post-harvest losses claim an estimated 35–50% of harvested produce in sub-Saharan Africa [1]. In Uganda, maize losses alone reach hundreds of thousands of tonnes annually, with aflatoxin, which is a serious storage-related health hazard. A lifecycle-level diagnosis shows that these losses stem less from a scarcity of documented knowledge than from information-systems failures [2]. Expert post-harvest knowledge is fragmented and statically stored; retrieval is keyword-bound and cannot bridge the semantic gap between farmer queries and institutionally encoded guidance; dissemination is one-way, and existing LLM/RAG agricultural advisors neglect the knowledge-intensive post-harvest phase and low-resource languages [3, 4]. Following a design science research approach, in which the designed artifact and its justified decisions constitute the contribution, we present the design and in-progress evaluation of a deployed multilingual post-harvest-specific advisory platform that grounds advice in verified institutional knowledge and answers farmers’ questions in their languages. 2 System Design and Key Decisions The system implements a multi-stage retrieval-augmented pipeline in which each design decision targets a specific lifecycle failure. (1) A structured knowledge base curated from verified institutional post-harvest sources replace fragmented, static repositories and supply the grounding corpus for retrieval. (2) Semantic retrieval over this corpus replaces keyword matching, resolving the query-document vocabulary mismatch. (3) An English-pivot translate-and-retrieve design lets queries in local languages like Luganda reach an English knowledge base through a translation backbone, circumventing the scarcity of native-language retrieval corpora. (4) A confidence gate withholds or redirects low-confidence responses rather than emitting unsupported advice, a safety-critical choice where hallucinated guidance has real consequences. (5) At the generation stage, the system supports two interchangeable configurations whose relative value is an open question: a prompt-engineered base model and the same base model with a lightweight QLoRA adapter (4-bit; small open models of 1.7B-4B parameters). Both operate on identical retrieved context, isolating the effect of parameter-efficient fine-tuning at generation while keeping deployment within commodity hardware limits. (6) A conversational multilingual interface delivers context-sensitive, decision-supporting responses in place of one-way broadcast. 3 Evaluation Design The evaluation, designed as part of the artifact because reliability rather than fluency determines value here, centers on a held-out benchmark, Agri-Advisor-QA: 358 expert-validated question-answer items across four staple crops (maize, beans, groundnuts, and sorghum) and six post-harvest themes, including dedicated hallucination-stress and out-of-scope redirect strata that probe factual grounding and safe refusal. Items are validated by a domain expert, completed in English before translation into Luganda, and the benchmark is held out from all model training. The protocol combines BERTScore and chrF, Recall@k, a five-dimension expert accuracy rubric (factual correctness and safety), and the System Usability Scale, applied across the different languages and to both generation configurations (prompted base vs. QLoRA-adapted) with retrieval held constant. The system is currently in production, and this evaluation is underway. 4 Conclusion We contribute a design: a reference architecture with a matched evaluation protocol, for retrieval-grounded, multilingual, decision-supporting post-harvest advisory in a low-resource setting, together with the rationale linking each design decision to a documented lifecycle failure. A central evaluated question is whether parameter-efficient adaptation improves grounded factual reliability over a strong prompted baseline within the same retrieval pipeline. The artifact shows how small, grounded models can move agricultural advisory beyond broadcast delivery toward context-sensitive decision support, offering a template transferable to other low-resource, high-stakes domains. References. [1] Ariong, R.M., Okello, D.M., Otim, M.H., Paparu, P. The cost of inadequate postharvest management of pulse grain: farmer losses due to handling and storage practices in Uganda. Agriculture & Food Security 12(1):20, 2023. [2] Alavi, M., Leidner, D.E. Knowledge management and knowledge management systems: conceptual foundations and research issues. MIS Quarterly 25(1):107–136, 2001. [3] Mapiye, O., et al. Information and communication technologies (ICTs) for disseminating agricultural information and services to smallholder farmers in sub-Saharan Africa. Information Development 39(3):638–658, 2023. [4] Singh, N., et al. Farmer.Chat: scaling AI-powered agricultural services. arXiv:2409.08916 (2024).