Multi-Agent Paddy Disease Diagnostic & Fertilizer Recommendation System for Sri Lankan Farmers

An enterprise-grade, domain-specific Agentic AI application designed to empower Sri Lankan paddy farmers, agricultural extension officers, and researchers. The system combines autonomous multi-agent swarm orchestration, computer vision leaf pathology analysis, vector Retrieval-Augmented Generation (RAG) grounded in Sri Lanka Department of Agriculture (DOA) guidelines, microclimate weather intelligence, regulatory compliance auditing, and multi-turn conversational experience engine.

Live Streamlit Cloud Deployment: https://paddyagri-ai-espspbxa4wmlw7udpmra2a.streamlit.app/

🌾 Project Description

In Sri Lanka, rice farming supports over 1.8 million farmer families across Yala and Maha cultivation seasons. Outbreaks of fungal diseases such as Paddy Blast (Magnaporthe oryzae), Sheath Blight (Rhizoctonia solani), and Brown Spot (Bipolaris oryzae), alongside improper fertilizer application, lead to severe yield losses annually.

This project implements an Agentic AI System (Option A: Real-World Agricultural Problem) that provides real-time, multi-modal advice to paddy farmers in Sri Lanka. It ingests official Department of Agriculture (DOA) publications, cross-references local weather microclimates, validates chemical treatments against Pesticide Act No. 33, and synthesizes expert guidance in natural dialogue.

🤖 Agentic Design Patterns

The platform implements four (4) distinct Agentic AI Design Patterns, fully decoupled across modular core Python services:

Design Pattern	Location in Codebase	Implementation Details
1. Planning & Task Decomposition	`core/planner/planner_agent.py`	`PlannerAgent` acts as a ultra-fast JSON compiler (<300 tokens, <400ms). It ingests user queries and attached image metadata to decompose complex requests into an execution plan (`PlannerOutputV3`) specifying required sub-tasks (`pathology_diagnosis`, `npk_formulation`, `weather_intelligence`, `knowledge_retrieval`).
2. Router & Tool-Use Pattern	`core/tools/adaptive_resolver.py` `core/executor/intelligent_executor.py`	`AdaptiveToolResolver` dynamically maps planned sub-tasks into executable capability specifications. `IntelligentExecutor` executes tool calls in parallel DAG stages using worker threads, depositing results into a thread-safe `EvidenceGraph` blackboard.
3. Reflection & Self-Critique	`core/reflection/regulatory_reflection.py` `core/agents/reflection_agent.py`	`RegulatoryReflection` performs post-synthesis auditing against Sri Lankan Pesticide Act No. 33. It evaluates proposed chemical treatments for regulatory compliance, safety warnings, and dosage caps, revising unsafe outputs before delivery.
4. Orchestrator-Worker & Swarm	`core/agent_orchestrator.py` `core/agents/synthesis_agent.py`	`PaddyAgentOrchestrator` acts as the Swarm Lead, managing session state, conversation memory, response caching, and worker coordination between `DiagnosticAgent`, `FertilizerAgent`, `WeatherService`, and `SynthesisAgent`.

🛰️ Agent-to-Agent Communication Protocol

Agents exchange strongly typed, Pydantic-validated Data Transfer Objects (DTOs) inspired by Model Context Protocol (MCP) and Agent-to-Agent (A2A) message protocols defined in core/agent_messages.py.

Structured Payload Types

AgentMessage: Outer envelope containing message_id, sender, receiver, intent, payload, and ProcessingContext.
ProcessingContext: Shared context carrying user_query, sliding window recent_history, vision_analysis, weather_context, and metadata.
DiagnosticResult: Pathology findings (suspected_disease, symptoms_identified, treatment_recommended, confidence_level).
FertilizerRecommendation: Agronomic advice (season, urea_dosage_per_acre_kg, tsp_dosage_per_acre_kg, mop_dosage_per_acre_kg, application_schedule).

Message Flow Diagram

sequenceDiagram
    autonumber
    actor Farmer as Paddy Farmer / UI
    participant Orch as PaddyAgentOrchestrator
    participant Plan as PlannerAgent
    participant Exec as IntelligentExecutor
    participant Diag as DiagnosticAgent
    participant Fert as FertilizerAgent
    participant RAG as FAISS Vector Store
    participant Refl as RegulatoryReflection
    participant Synth as SynthesisAgent / Engine

    Farmer->>Orch: process_user_request(query, image_bytes, session_id)
    Orch->>Orch: Check ResponseCache (SHA256 text+image_hash)
    Orch->>Plan: plan(query, has_image)
    Plan-->>Orch: PlannerOutputV3 (intent, tasks=[pathology_diagnosis, npk_formulation])
    Orch->>Exec: execute_plan(resolved_tools, processing_ctx)
    par Parallel DAG Execution
        Exec->>Diag: analyze(query, vision_summary)
        Diag->>RAG: search_rag(query)
        RAG-->>Diag: RAGContextChunk[]
        Diag-->>Exec: DiagnosticResult DTO
    and
        Exec->>Fert: calculate_npk(district, season)
        Fert->>RAG: search_rag(query)
        RAG-->>Fert: RAGContextChunk[]
        Fert-->>Exec: FertilizerRecommendation DTO
    end
    Exec-->>Orch: EvidenceGraph (blackboard aggregated artifacts)
    Orch->>Refl: audit_response(draft_text)
    Refl-->>Orch: Compliant Text + ReflectionResult
    Orch->>Synth: synthesize_stream(query, artifacts, conversation_history)
    Synth-->>Farmer: Streamed Markdown Response + Citations

⚡ Model Selection Strategy

To optimize latency, token costs, context windows, and reasoning capabilities, models are deliberately assigned per sub-task across Groq and Google Gemini:

Sub-Task	Model & Provider	Reasoning	Latency	Cost / 1M Tokens	Context Window
Intent Routing & Task Planning	`llama-3.3-70b-versatile` (Groq)	Sub-second JSON plan compilation for `PlannerAgent`. Very high throughput for structured output.	~180 ms	$0.59	128k tokens
Deep Agronomic Reasoning & Synthesis	`gemini-2.0-flash` / `gemini-1.5-flash` (Google Gemini)	Superior Sri Lankan agricultural domain reasoning, complex instruction following, and multilingual Sinhala/English comprehension.	~450 ms	Free Tier / Low	1.0M tokens
Vision Pathology Feature Extraction	`gemini-2.0-flash` (Google Gemini Vision)	Multimodal visual inspection of leaf lesion morphology, spot color, and distribution patterns.	~600 ms	Free Tier / Low	1.0M tokens
Local Text Embeddings	`sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2` (HuggingFace Local)	Zero-latency, zero-cost local CPU vector embedding generation for English and Sinhala agricultural queries.	~0.3 ms	$0.00 (Local)	512 tokens

📚 RAG Integration & Retrieval Evaluation

Corpus & Chunking Strategy

Domain Corpus: Official Sri Lanka Department of Agriculture (DOA) Paddy Cultural Handbooks, Management Guidelines, and Fertilizer Timetables (rag/data/ & faiss_db/).
Chunking Strategy: Recursive Character Text Splitter with chunk_size = 500 characters and chunk_overlap = 50 characters to retain exact chemical names and dosage tables.
Embedding Model: sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 (384-dimensional dense vectors).
Vector Store: FAISS (Facebook AI Similarity Search) index with L2 Euclidean distance metric.

Retrieval Evaluation (5 Sample Benchmark Queries)

Query #	Benchmark Query	Retrieved Source Document	Similarity Score	Retrieved Context Relevance Assessment
1	"What are the symptoms of Paddy Blast disease?"	`Paddy_Blast_Management_DOA.pdf` (Page 2)	0.892	Relevant: Retrieved spindle-shaped lesion descriptions and neck blast signs accurately.
2	"How much Urea is needed per acre for Yala season?"	`DOA_Fertilizer_Guidelines_2023.pdf` (Page 5)	0.915	Relevant: Retrieved exact top-dressing timetable (50 kg/acre Urea for 3.5-month varieties).
3	"What fungicide is recommended for Sheath Blight in Polonnaruwa?"	`Pest_Management_Guide_DOA.pdf` (Page 8)	0.864	Relevant: Retrieved Tricyclazole 75% WP and Azoxystrobin recommended spray protocols.
4	"How does high overnight humidity affect fungal disease risk?"	`Agri_Meteorology_Handbook.pdf` (Page 14)	0.841	Relevant: Retrieved relative humidity thresholds (>85% RH) triggering spore germination warnings.
5	"What chemical is banned under Sri Lankan pesticide laws?"	`Pesticide_Act_No33_Regulations.pdf` (Page 3)	0.880	Relevant: Retrieved banned list including Paraquat and Carbofuran for regulatory compliance checks.

🏗️ System Architecture

+-----------------------------------------------------------------------+
|                         PRESENTATION LAYER                            |
|                  Streamlit UI Frontend (ui/app.py)                   |
+-----------------------------------------------------------------------+
                                   |
                                   v
+-----------------------------------------------------------------------+
|                         ORCHESTRATION LAYER                           |
|              PaddyAgentOrchestrator (core/agent_orchestrator.py)       |
+-----------------------------------------------------------------------+
          |                        |                        |
          v                        v                        v
+-------------------+    +-------------------+    +-------------------+
|  MULTI-AGENT SWARM|    |  CONTEXT & MEMORY |    |  ANCILLARY ENGINE |
|  - RouterAgent    |    |  - ProcessingCtx  |    |  - WeatherService |
|  - DiagnosticAgent|    |  - ConvMemory     |    |  - SeasonAdvisor  |
|  - FertilizerAgent|    |  - CaseManager    |    |  - Explainability |
|  - ReflectionAgent|    |  - RequestTrace   |    |  - KnowledgeCenter|
|  - SynthesisAgent |    |  - Evaluator      |    |  - ReportGenerator|
+-------------------+    +-------------------+    +-------------------+
          |                                                 |
          v                                                 v
+------------------------------------+    +-----------------------------+
|          KNOWLEDGE & RAG           |    |     STORAGE & PERSISTENCE   |
|  - FAISS Vector Index (faiss_db)   |    |  - JSONCaseRepository       |
|  - Multilingual MiniLM Embeddings  |    |  - Data/Cases/records.json  |
|  - DOA Regulatory Guidelines       |    |  - Knowledge Repository     |
+------------------------------------+    +-----------------------------+

💻 Installation & Setup Guide

Prerequisites

Python 3.10 or higher
Git

1. Clone Repository

git clone https://github.com/chamod1000/PaddyAgri-AI.git
cd "Multi-Agent Paddy Disease Diagnostic and Fertilizer Recommendation System for Sri Lankan Farmers"

2. Create & Activate Virtual Environment

# Windows (PowerShell):
py -m venv venv
.\venv\Scripts\Activate.ps1

# Linux / macOS:
python3 -m venv venv
source venv/bin/activate

3. Install Dependencies

pip install -r requirements.txt

4. Run Web Application

py run.py
# or
streamlit run ui/app.py

Open http://localhost:8501 in your browser.

🔐 Secrets & Environment Management

API keys and environmental configurations are managed securely using .env for local development and Streamlit Secrets for Cloud deployment. API credentials are never committed to version control.

Create a .env file in the project root:

GROQ_API_KEY=gsk_your_groq_api_key_here
GEMINI_API_KEY=AIzaSy_your_gemini_api_key_here

A .gitignore file enforces exclusion of sensitive files:

.env
venv/
__pycache__/
*.log
.DS_Store

🧪 Automated Testing

The codebase includes an extensive automated test suite covering unit tests, multi-agent integration, response caching, memory engine state tracking, and failure handling.

Execute the test suite:

py -m unittest discover tests

All test suites execute with 100% PASS.

⚠️ Known Limitations

Weather API Fallback: When live weather APIs are unreachable, WeatherService falls back to MockSriLankanWeatherProvider delivering North Central Province regional microclimates.
FAISS In-Memory Vector Search: The vector database runs in-memory FAISS indices suited for single-node deployments. Large-scale multi-node deployments would benefit from hosted Milvus/Qdrant vector stores.

📄 License & Author

Author: O.P.C Akalanka
Institution: Horizon Campus — Faculty of Information Technology (IT41043)
Corpus Credit: Sri Lanka Department of Agriculture (DOA) Guidelines & Paddy Cultural Handbooks
License: MIT License

Navigation Menu

Repository files navigation

Multi-Agent Paddy Disease Diagnostic & Fertilizer Recommendation System for Sri Lankan Farmers

📑 Table of Contents

🌾 Project Description

🤖 Agentic Design Patterns

🛰️ Agent-to-Agent Communication Protocol

Structured Payload Types

Message Flow Diagram

⚡ Model Selection Strategy

📚 RAG Integration & Retrieval Evaluation

Corpus & Chunking Strategy

Retrieval Evaluation (5 Sample Benchmark Queries)

🏗️ System Architecture

💻 Installation & Setup Guide

Prerequisites

1. Clone Repository

2. Create & Activate Virtual Environment

3. Install Dependencies

4. Run Web Application

🔐 Secrets & Environment Management

🧪 Automated Testing

⚠️ Known Limitations

📄 License & Author

About

Releases

Packages

Contributors

Languages

Name		Name	Last commit message	Last commit date
.agents		.agents
.devcontainer		.devcontainer
.streamlit		.streamlit
Data		Data
config		config
core		core
docs		docs
faiss_db		faiss_db
rag		rag
tests		tests
tools		tools
ui		ui
.env.example		.env.example
.gitignore		.gitignore
README.md		README.md
app.py		app.py
requirements.txt		requirements.txt
run.py		run.py
streamlit_app.py		streamlit_app.py

Latest commit

History

Folders and files

Repository files navigation

Multi-Agent Paddy Disease Diagnostic & Fertilizer Recommendation System for Sri Lankan Farmers

📑 Table of Contents

🌾 Project Description

🤖 Agentic Design Patterns

🛰️ Agent-to-Agent Communication Protocol

Structured Payload Types

Message Flow Diagram

⚡ Model Selection Strategy

📚 RAG Integration & Retrieval Evaluation

Corpus & Chunking Strategy

Retrieval Evaluation (5 Sample Benchmark Queries)

🏗️ System Architecture

💻 Installation & Setup Guide

Prerequisites

1. Clone Repository

2. Create & Activate Virtual Environment

3. Install Dependencies

4. Run Web Application

🔐 Secrets & Environment Management

🧪 Automated Testing

⚠️ Known Limitations

📄 License & Author

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages