pitsophinnias/Mosupisi-An-AI-Powered-Agricultural-Extension-Agent-for-Smallholder-Farmers-in-Lesotho

AI-powered agricultural advisory platform for smallholder farmers in Lesotho. Delivers bilingual (English + Sesotho) guidance on planting, pest control, and weather-aware decisions using a local quantized LLM and RAG - offline-first, with SMS alerts and a React PWA frontend. # Mosupisi - AI-Powered Agricultural Extension Platform **Mosupisi** (Sesotho for "guide" or "one who shows the way") is an AI-powered agricultural advisory platform designed specifically for smallholder farmers and extension officers in **Lesotho**. It delivers practical, bilingual (English + Sesotho) guidance on planting, pest control, weather-aware decisions, and general farming queries using a **local quantized LLM** and **Retrieval-Augmented Generation (RAG)**. - **Offline-first**: works in low-connectivity rural areas with data synchronisation when connectivity is restored. - **Locally grounded**: RAG knowledge base built from official Lesotho Meteorological Services (LMS) Agrometeorological Bulletins (2010-2026). - **Bilingual**: full English and Sesotho interface with on-the-fly translation of AI advice. - **Proactive alerts**: weather hazards, crop milestones, pest risks, and spray windows via SMS and push notifications. --- ## Table of Contents - Architecture Overview - Model Training - Services & Ports - Prerequisites - Getting Started - 1. Clone the Repository - 2. Environment Variables - 3. Model Placement (Critical) - 4. Install Dependencies - Running Mosupisi - Option A: Automated Scripts (Recommended) - Option B: Manual Start (Order matters) - Health Checks - Using the Platform - Troubleshooting - Next Steps & Future Work --- ## Architecture Overview Mosupisi follows a **microservices architecture** with independent, loosely coupled services. A central **LLM service** loads the quantised Mistral model (`mosupisi-q4.gguf`) and serves all other modules. RAG retrieval is performed via ChromaDB using sentence-transformer embeddings (`all-MiniLM-L6-v2`). ### How the System Fits Together At the centre of the platform is the **LLM Service** (port 3004), which loads the quantised Mistral model (`mosupisi-q4.gguf`) and exposes a single `/infer` endpoint. Every other backend service that needs AI-generated text calls this endpoint - none of them lo …