We introduce PestIDBot, an integrated platform for agricultural pest manage- ment that combines rigorously tested deep-learning models and a retrieval- augmented conversational agent to address challenges in pest identification and information retrieval. The pest classifiers, InsectNet and WeedNet, are trained on large-scale datasets and incorporate out-of-distribution detection and conformal prediction to enhance reliability. The conversational agent,AgLLM, utilizes large language models and has been thoroughly evaluated using retrieval-augmented generation metrics and expert assessments to ensure accurate and contextually relevant responses. By integrating these components, PestIDBot provides a comprehensive tool that delivers reli- able pest identification and information retrieval within a unified interface.
PestIDBot interface showing pest identification result and chatbot interaction.
Illustration of the conversational retrieval chain used in the AgLLM chatbot.
Retrieval Performance without Metadata Filtering (Adjust the y-axes in RAG plots, remove precision and just merge both of them)
Retrieval Performance with Metadata Filtering
Distribution of generated responses for (a) Insect-related and (b) Weed-related research areas. The inner ring represents the main categories, while the outer ring shows the split between Knowledge (diagonal lines) and Management (dots) within each category.