KrishokChat
Safety-aware Bengali agricultural advisory RAG with crop vision disease diagnostics.
1. Problem Context & Objectives
Smallholder farmers in Bangladesh lack accessible, dialect-aware advisory systems and face catastrophic crop losses from incorrect pesticide recommendations and chemical hallucination.
2. Technical Architecture & Implementation
Engineered a 4-stage agent pipeline: Dialect normalizer → Dual-stage retriever (BM25 k1=2.2/b=0.4 + FAISS mE5-small) → Chemical safety refusal audit sink → Grounded Bengali generation. Includes a multi-class vision pipeline for crop leaf diagnostics.
3. Outcomes, Benchmarks & Empirical Validation
Evaluated across 2,120 knowledge nodes and 20,112 safety evaluation records with provenance verification, achieving 88% wheat disease accuracy and 99.8% crop library coverage.
Source Code Inspection & Local Replication
This system is maintained in an open-source repository with full git commit history, environment configuration templates (.env.example), and dependency manifests.