MODULAR RETRIEVAL-AUGMENTED GENERATION FOR AGENTIC TOBACCO CROP DISEASE DIAGNOSIS AND SOIL ADVISORY

Abstract

In spite of the fact that the tobacco farmers using mallholders need timely evidence-based guidance for tackling the problems of leaf diseases and soil fertility, the agentic crop-advisory systems usually combine perception agents (classifiers based on image and sensor data) and a hardcoded rule table, but not a genuine knowledge retrieval system. In this paper, we redesign one of these prototypes—an agentic AI system for tobacco called "Tobacco Agentic AI," which consists of a soil condition agent, leaf disease agent, and recommendation agent—using a modular RAG framework that breaks down the lookup process into five distinct steps. For the evaluation of our method, we compile two benchmark datasets with 10,000 records each: (i) Tobacco Leaf Disease Image subsample – a class-balanced subset derived from the publicly available Kaggle tobacco lesions dataset compiled by Chen et al., and (ii) Kaggle Crops-NPK soil fertility dataset 10,000 record stratified subsample. Using these datasets, we generate a set of 200 queries with disease-soil evaluation tasks and tag-based ground truth. We evaluate three setups: a baseline Naive RAG (single vector top-k search), Modular RAG without query expansion and template/LLM generation (retrieval only), and full Modular RAG (query expansion, hybrid TF-IDF and metadata filtering retrieval, exact match re-ranking, template/LLM generation). The latter setup yields an increase in Precision@3 from 0.61 to 0.84, Recall@3 from 0.70 to 1.00, and Reciprocal Rank from 0.855 to 1.000 compared to the Naive RAG baseline while incurring a negligibly small latency overhead (0.79 ms vs 0.85 ms per query on commodity CPU hardware). Compared to the simple baseline, the whole pipeline improves average Precision@3 from 0.61 to 0.84, Recall@3 from 0.70 to 1.00, and average Reciprocal Rank from 0.855 to 1.000 at a slight cost of latency (0.79 ms vs. 0.85 ms per query on commodity CPU hardware). We also benchmark our system against two similar but not directly related works published during the years 2023-2026 – an advice framework based on the RAG multi-crop model for LLMs and deep-learning lesion segmentation (MD-Unet) in tobacco plants. We also discuss architecture limitations and future extensions.