Image (512×512 plant photo)
    ↓ CLIP ViT-B/32
    [512-dim visual embedding]
        +
Sensor Window (pH, EC, temp, humidity, VPD, biomass, ...)
    ↓ SensorEncoder (LSTM + linear projection)
    [7-dim sensor embedding]
        =
FMU Vector [519-dim]  ←→  stored in Qdrant with full metadata payload

┌─────────────────────────────────────────────────────────────────┐
│                     DEMETER AGENT CYCLE                         │
│                                                                 │
│  1. FetchingAgent                                               │
│     └─ Polls simulator/ADT → builds FMU → stores in Qdrant      │
│                                                                 │
│  2. JudgeAgent  (LangGraph)                                     │
│     └─ Retrieves N-1 FMU → runs Azure CV (visual forensics)     │
│     └─ Queries FarmMemory (Mem0) → deliberates via LLM          │
│     └─ Files reward score → trains Bandit                       │
│                                                                 │
│  3. SupervisorAgent → ContextualBandit                          │
│     └─ Encodes current FMU as context vector (519-dim)          │
│     └─ Bandit selects 1 of 15 strategies (LinGreedy)            │
│     └─ Strategy instruction passed to expert agents             │
│                                                                 │
│  4. ResearcherAgent                                             │
│     └─ Queries Knowledge_Base (RAG) for crop-specific data      │
│                                                                 │
│  5. AtmosphericAgent  (LangGraph)                               │
│     └─ Tools: ask_rag, web_search, calculate_vpd, diagnose_plant│
│     └─ Generates climate action plan (temp/humidity/CO2/light)  │
│     └─ Simulates plan → retry loop if unsafe                    │
│                                                                 │
│  6. WaterAgent  (LangGraph)                                     │
│     └─ Tools: ask_rag, web_search, calculate_vpd, diagnose_plant│
│     └─ Generates nutrient plan (pH dosing, EC adjustment)       │
│     └─ Simulates plan → retry loop if unsafe                    │
│                                                                 │
│  7. SupervisorAgent → synthesize_plan                           │
│     └─ Merges atmospheric + water plans                         │
│     └─ Detects conflicts, enforces safety bounds                │
│     └─ Dispatches final FarmAction to Simulator/ADT             │
└─────────────────────────────────────────────────────────────────┘

crop_id              String (unique, required)
crop                 String
stage                String
sequence_number      Number
cycle_duration_hours Number  (auto-set per crop type: lettuce/basil=1h, tomato/strawberry=2h)
total_crop_lifetime_days Number
simulated_age_hours  Number
planted_at           Date
last_updated         Date
sensors              { pH: [Number], EC: [Number], temp: [Number], humidity: [Number] }
sensor_ids           { ph_sensor, ec_sensor, temp_sensor, humidity_sensor }
location             String
notes                String
image_url            String
action_taken         Mixed
outcome              String
explanation_log      String
bandit_action_id     Number
strategic_intent     String
reward_score         Number
visual_diagnosis     String
schema_version       String  (current: "1.2")

demeter/
├── agent/
│   ├── main_agent.py                # Orchestration loop (runs the 7-agent cycle)
│   ├── memory.py                    # FarmMemory class (Mem0 + Qdrant)
│   ├── guardrails/
│   │   ├── README.md                # Guardrails documentation
│   │   └── validation.py            # Input sanitization, bounds checking, injection detection
│   ├── model/
│   │   └── plant_disease_model.pt   # Plant disease detection model
│   ├── Marl/
│   │   ├── bandit.py                # Contextual Bandit (LinGreedy, 15 arms)
│   │   ├── strategies.py            # Strategy definitions
│   │   ├── train-bandit.py          # Offline training script
│   │   └── model_bandit_greedy.pkl  # Trained bandit model weights
│   ├── Sentinel/
│   │   ├── agent.py                 # FMUBuilder - creates fused vectors
│   │   ├── fmu.py                   # FMU dataclass
│   │   └── Encoders/
│   │       ├── Vision.py            # CLIP ViT-B/32 image encoder
│   │       └── TimeSeries.py        # LSTM-based sensor encoder
│   ├── Qdrant/
│   │   ├── Client.py                # Qdrant client singleton
│   │   ├── Setup.py                 # Collection initialization
│   │   ├── Store.py                 # FMU storage helpers
│   │   └── Search.py                # Similarity search helpers
│   ├── sub_agents/
│   │   ├── base_agent.py            # BaseReasoningAgent (Azure OpenAI client)
│   │   ├── fetching_agent.py        # FetchingAgent - polls simulator, builds FMU
│   │   ├── judge_agent.py           # JudgeAgent - reward evaluation (LangGraph)
│   │   ├── atmospheric_agent.py     # AtmosphericAgent - climate planning (LangGraph)
│   │   ├── water_agent.py           # WaterAgent - nutrient planning (LangGraph)
│   │   ├── Supervisor.py            # SupervisorAgent - strategy + synthesis
│   │   ├── Researcher.py            # ResearcherAgent - RAG knowledge retrieval
│   │   ├── Explainer.py             # ExplainerAgent - chain-of-thought log generation
│   │   ├── Doctor.py                # VisionAgent - plant disease detection (Azure Custom Vision)
│   │   └── water_and_atmospheric_dependencies/
│   │       ├── state.py             # LangGraph AgentState definition
│   │       ├── nodes.py             # LangGraph node functions
│   │       ├── tools.py             # LangChain tools (calculate_vpd, web_search)
│   │       ├── physics_engine.py    # LLM-based plan safety simulator
│   │       └── retrieval.py         # LangChain tools (ask_rag, diagnose_plant, ask_memory)
│   └── tools/
│       ├── actuation.py             # Actuator command builders
│       ├── db_tools.py              # Database utility tools
│       ├── processing_tools.py      # Sensor processing utilities
│       └── reset_memory.py          # Memory reset utility
│
├── backend/
│   ├── server/
│   │   ├── main.py                  # FastAPI server (Python) - agent HTTP endpoints
│   │   ├── functions.py             # Backend utility functions
│   │   ├── rag_brain.py             # PDF ingestion pipeline for Knowledge Base
│   │   ├── create-index.py          # Qdrant index setup script
│   │   └── reset-db.py              # Database reset utility
│   └── node_server/
│       ├── index.js                 # Express.js server - crop CRUD API + MongoDB init
│       ├── routes/
│       │   ├── farmRoutes.js        # Farm/Qdrant history routes
│       │   └── cropRoutes.js        # Crop CRUD routes
│       ├── controllers/
│       │   ├── farmController.js
│       │   └── cropController.js
│       ├── schema/
│       │   └── cropSchema.js        # Mongoose schema (v1.2)
│       └── config/db.js             # MongoDB + legacy DB connection
│
├── frontend/
│   ├── src/
│   │   ├── App.js                   # Router, providers, onboarding gate
│   │   ├── pages/                   # All 10 page components
│   │   ├── components/              # Sidebar, AgentWidgets, Onboarding
│   │   ├── hooks/                   # useFarmData, useSettings, useTranslation
│   │   ├── api/
│   │   │   ├── agentApi.js
│   │   │   └── farmApi.jsx          # MongoDB CRUD calls
│   │   ├── utils/                   # translations.js, dataUtils.js
│   │   └── data/mockData.js         # Mock data for testing
│   └── tailwind.config.js
│
├── simulator/
│   └── main.py                      # Multi-batch DigitalTwin fleet + Azure ADT sync
│
├── yolov8n.pt                       # Pre-trained YOLOv8 model
├── Knowledge_Base/                  # Drop agronomic PDFs here for RAG ingestion
├── requirements.txt
└── README.md

git clone https://github.com/maydayv7/demeter.git
cd demeter
pip install -r requirements.txt

# Azure OpenAI
AZURE_OPENAI_API_KEY=your_azure_openai_key
AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com/
AZURE_OPENAI_DEPLOYMENT_NAME=gpt-4.1
AZURE_OPENAI_API_VERSION=2024-12-01-preview

# Qdrant
QDRANT_URL=http://localhost:6333
QDRANT_API_KEY=your_qdrant_api_key   # Only needed for Qdrant Cloud

# Azure Custom Vision
AZURE_API_KEY=your_azure_api_key
AZURE_ENDPOINT=your_azure_endpoint
AZURE_PROJECT_ID=your_azure_project_id
AZURE_PREDICTION_KEY=your_azure_prediction_key
AZURE_URL=your_azure_url
AZURE_ITERATION_NAME=DemeterDoctor
DATASET_FOLDER="train"

# Simulator
SIMULATOR_STATE_URL=http://localhost:8001/simulation/state
SIMULATOR_ACTION_URL=http://localhost:8001/simulation/action
SIMULATOR_PORT=8001

# Azure Digital Twins
ADT_URL=your-adt-instance.digitaltwins.azure.net

# Backend
PORT=3001
MONGODB_URI=your_mongodb_connection_string

# Frontend
REACT_APP_AGENT_API_URL=http://localhost:8000
REACT_APP_FARM_API_URL=http://localhost:3001/api

docker run -p 6333:6333 -p 6334:6334 qdrant/qdrant

python backend/server/create-index.py

python backend/server/rag_brain.py

python simulator/main.py
# Runs on http://localhost:8001

# Python FastAPI server
python backend/server/main.py

# NodeJS Express server
cd backend/node_server
npm install && npm start

cd frontend
npm install
npm start
# Runs on http://localhost:3000

python agent/main_agent.py