The literature behind AgroVision.

A perspective on Multi-Agent Systems powered by LLMs for the geosciences. Presents PANGAEA GPT, a specialised MAS pipeline integrated with the PANGAEA Earth-and-Environmental-Science database, demonstrating how MAS-driven workflows manage complex, multi-format scientific data.

Cited to justify the multi-agent architecture. Agricultural decision support spans crops, soil, weather, and pests — exactly the kind of cross-domain workload single-agent setups handle poorly.

A modular, role-specialised multi-agent framework for smallholder farmers. Uses a chain-of-responsibility pattern to coordinate autonomous agents with real-time tools (weather, geospatial data) and a dedicated critique agent for hallucination control. Supports multilingual interaction.

Closest peer to AgroVision's architecture. Informs the orchestrator + specialised-agent + judge pattern that AgroVision uses.

Soil2Harvest-AI — an agentic framework for end-to-end crop yield optimisation. Combines soil image classification (Custom CNN, 92.88% accuracy), regression-based soil parameter estimation, real-time weather integration, and explainable-AI overlays (SHAP, LIME).

Validates the model of orchestrating multiple agricultural ML sub-tasks behind a single user-facing decision support interface.

An agent-based model combining fuzzy logic and system dynamics to evaluate sugarcane vulnerability to climate change. Uses a three-stage methodology: vulnerability indicators, crop vulnerability, and total system vulnerability.

Demonstrates how agent-based decision support translates climate variables into actionable crop-vulnerability assessments — directly applicable to AgroVision's climate-coupled reasoning.

A multi-agent system that operationalises the Sustainable Agricultural Intensification framework. Pairs an Environmental Module (nutrient inputs and outputs) with a Behavioural Module (farmer decisions on fertilizer, manure, biological N fixation) to simulate sustainability trajectories on managed plots.

Reference for separating environmental modelling from behavioural decision-making — informs AgroVision's split between Weather/Soil agents and the orchestrator/advisory layer.

A RAG-based agricultural advisory system covering crop cultivation, pest and disease management, and fertilizer application across five crops (maize, ragi, sweet potato, cotton, groundnut). Processes Package of Practices documents via semantic chunking and benchmarks four LLMs.

Concrete evidence that RAG on government and extension-service documents produces deployable farmer advisory — supports AgroVision's choice to ground responses in verified agronomic corpora.

A configurable RAG framework that decomposes information access into declarative stages (query refinement, sub-query planning, retrieval, synthesis, evaluation). Integrates 1B-parameter LLMs with adaptive retrievers and enforces deterministic citation (59–73% perfect citation accuracy) to eliminate citation hallucination.

Architectural reference for AgroVision's Agentic RAG layer — particularly the deterministic citation mechanism, which aligns with the Judge Agent and the project's 'no fabricated claims' stance.