Prompt We are starting a completely new implementation of an AI-powered crop disease diagnosis system called LeafDoctor AI. Forget all previous project architecture, interfaces, factories, dependency injection, repositories, and service layers. We are building a WORKING MVP first. After the MVP is complete, we will refactor into a production architecture. ================================================== PROJECT GOAL ================================================== Implement a complete working pipeline. Farmer uploads leaf image ↓ Image is stored inside uploads/ ↓ Google Gemini Vision diagnoses the image ↓ Gemini returns - Crop - Problem Label - Confidence ↓ Lookup PostgreSQL database ↓ Retrieve - Class_ID - Treatment - Chemical - Dosage - SMS - etc. ↓ Store diagnosis history ↓ Return JSON response ================================================== IMPORTANT BUSINESS RULES ================================================== Gemini is ONLY responsible for diagnosis. Gemini NEVER generates - recommendation - treatment - dosage - chemical - SMS Those always come from PostgreSQL. ================================================== AVAILABLE DATA ================================================== I already have these datasets. 1. crop_protection_labels.csv Contains Class_ID Label_Name Primary_Category This is the master label table. 2. normalized_class_treatment_mapping.csv Contains approximately 3090 records. Columns include Class_ID Master_Label_Name Crop Type of Problem Chemical & Formulation Chemical Type Revised Dosages SMS : Growth Stage 1 This is the recommendation database. These CSVs have already been generated by Gemini as a mapping layer. You should analyze these files before deciding how Gemini should respond. ================================================== VERY IMPORTANT ================================================== Do NOT assume Gemini should return Class_ID. First analyze the uploaded CSVs. If returning Label_Name is more reliable than returning Class_ID, then redesign the workflow accordingly. Accuracy is much more important than elegance. ================================================== DATABASE ================================================== Use PostgreSQL. Use SQLAlchemy ORM. Do NOT use SQLite. CSV/XLSX files will be imported only once. After import, PostgreSQL becomes the single source of truth. ================================================== IMAGE STORAGE ================================================== Every uploaded image must be stored inside uploads/ using a unique filename. Example uploads/20260708153022.jpg These images will later be used for training our own AI model. ================================================== FASTAPI ================================================== Create POST /diagnose Workflow 1. Receive uploaded image 2. Validate image 3. Save image inside uploads/ 4. Call Gemini Vision 5. Receive diagnosis 6. Lookup PostgreSQL 7. Store diagnosis history 8. Return JSON ================================================== PROJECT STRUCTURE ================================================== Keep everything SIMPLE. leafdoctor/ app.py config.py database.py models.py gemini.py service.py import_data.py uploads/ data/ requirements.txt .env No repositories. No interfaces. No dependency injection. No factories. No enterprise architecture. ================================================== DEVELOPMENT RULES ================================================== 1. Implement ONE file at a time. 2. Every file must be COMPLETE. 3. Never generate placeholders. 4. Never generate TODOs. 5. Never generate pseudo code. 6. Every file must run. 7. Wait for my confirmation before generating the next file. ================================================== VERY IMPORTANT INSTRUCTION ================================================== Before writing any code, carefully analyze the uploaded CSV/XLSX files. I want the MOST RELIABLE solution, not the most beautiful architecture. If you discover a better workflow than the one described above based on my datasets, explain it first before writing code. Do not blindly follow my assumptions if the data suggests a better approach. ================================================== FIRST TASK ================================================== First analyze all uploaded datasets and propose the final database schema and complete end-to-end workflow. Do not write any code yet. After I approve the workflow, start implementing the project beginning with database.py. leafdoctor/ │ ├── app.py ├── config.py ├── database.py ├── models.py ├── gemini.py ├── service.py ├── import_data.py ├── uploads/ ├── data/ ├── requirements.txt └── .env # LeafDoctor AI – MVP Development Context We are building a **completely new AI-powered crop disease diagnosis system** called **LeafDoctor AI**. This is a **working MVP**, not a production architecture. **Golden Rule** Do NOT redesign the architecture. Do NOT introduce Repository Pattern. Do NOT introduce Dependency Injection. Do NOT introduce Factories. Do NOT introduce Interfaces. Do NOT introduce Enterprise Architecture. The goal is to build a **working end-to-end system first**. We will refactor later after the MVP is complete. --- # Technology Stack Backend * Python 3.11+ * FastAPI * SQLAlchemy ORM * PostgreSQL * Google Gemini Vision (`google-genai`) * Pydantic v2 * pydantic-settings Image * Pillow CSV Import * pandas * openpyxl --- # Folder Structure ```text leafdoctor/ app.py config.py database.py models.py gemini.py service.py import_data.py import_utils.py uploads/ data/ requirements.txt .env ``` --- # Project Workflow Farmer uploads image ↓ Image stored inside uploads/ ↓ Gemini Vision diagnoses image ↓ Gemini returns ONLY * crop * label * confidence ↓ Backend validates result ↓ Lookup PostgreSQL ↓ Retrieve * Class_ID * Treatment * Chemical * Dosage * SMS ↓ Save diagnosis history ↓ Return JSON --- # Business Rule Gemini NEVER generates * treatment * dosage * recommendation * chemical * SMS Gemini ONLY performs diagnosis. Everything else comes from PostgreSQL. --- # Datasets ## Master Labels crop_protection_labels.csv Contains * Class_ID * Label_Name * Primary_Category Approximately 201 labels. This is the MASTER dataset. --- ## Recommendation Mapping normalized_class_treatment_mapping.csv Approximately 3090 records. Contains * Class_ID * Crop * Type of Problem * Chemical & Formulation * Chemical Type * Revised Dosages * SMS : Growth Stage 1 This is the recommendation dataset. --- # Data Analysis Findings We analyzed both datasets. Important findings 1. Master labels are unique. One Class_ID ↓ One Label 2. Treatment table contains mapping inconsistencies. Master_Label_Name is NOT reliable. There are shifted labels. Therefore DO NOT import Master_Label_Name. Always resolve labels using Class_ID ↓ protection_labels 3. PostgreSQL becomes the single source of truth. CSV files are used only once during import. --- # Database Design ## protection_labels id class_id label_name_en label_name_bn primary_category_en primary_category_bn created_at --- ## treatment_mapping id class_id crop_en crop_bn problem_type_en problem_type_bn chemical chemical_type dosage sms_en sms_bn created_at --- ## diagnosis_history id image_path crop_detected predicted_label class_id confidence treatment_id gemini_model raw_response verified_label verified_by verified_at created_at --- # Image Storage All uploaded images are stored. Example uploads/ 20260708153022_abcd1234.jpg These images will later be used to train our own AI model. --- # Import Strategy import_utils.py Responsibilities * Load CSV * Clean data * Validate duplicates * Detect orphan Class_ID * Detect label mismatch * Import PostgreSQL * Generate data/crop_label_mapping.json * Generate data/import_report.json --- # crop_label_mapping.json Generated during import. Example ```json { "Rice": { "labels": [ "Blast", "Brown Spot", "Leaf Folder" ], "class_ids": [ 12, 15, 18 ] } } ``` Gemini loads this JSON. NOT PostgreSQL. --- # Gemini Library google-genai Model gemini-2.5-flash Uses ONE API call. Not two-stage. Returns ```json { "crop":"Rice", "label":"Blast", "confidence":0.98 } ``` Uses response_schema with DiagnosisSchema NOT regex parsing. Validation * crop exists * label belongs to crop * confidence between 0 and 1 Returns DiagnosisResult dataclass. --- # Service Layer Workflow Save uploaded image ↓ Gemini diagnose ↓ Lookup ProtectionLabel ↓ Get Class_ID ↓ Lookup TreatmentMapping ↓ Save DiagnosisHistory ↓ Return JSON Important fixes already identified * Case-insensitive label lookup * Case-insensitive crop lookup * Rollback database transaction on failure * Ensure uploads directory exists * Remove singleton instance at bottom of service.py --- # App FastAPI Endpoints GET / GET /health GET /statistics GET /history GET /history/{id} DELETE /history/{id} POST /diagnose POST /diagnose Accepts multipart UploadFile. Stores image. Calls service. Returns diagnosis. --- # Configuration .env contains APP_NAME APP_VERSION DEBUG LOG_LEVEL DB_HOST DB_PORT DB_NAME DB_USER DB_PASSWORD GEMINI_API_KEY GEMINI_MODEL GEMINI_TEMPERATURE GEMINI_MAX_OUTPUT_TOKENS UPLOAD_DIR DATA_DIR MAX_UPLOAD_SIZE_MB config.py already exposes settings using pydantic-settings. --- # Current Project Status Completed ✓ database.py ✓ config.py ✓ models.py ✓ import_utils.py ✓ import_data.py ✓ gemini.py ✓ service.py ✓ app.py ✓ requirements.txt ✓ .env Some runtime fixes may still be required because the project has not yet been executed end-to-end. --- # Development Rule From now on Do NOT redesign anything. Do NOT rewrite entire files unless absolutely necessary. If an error occurs Fix ONLY that error. Keep all business logic unchanged. The objective is to finish the MVP as quickly as possible and stabilize it through testing. pip install -r requirements.txt playwright install python3 -m venv venv source venv/bin/activate pip install --upgrade pip pip install -r requirements.txt To run api for ssh session uvicorn api_filter:app --reload --port 8000 To run continuously nohup uvicorn land_api:app --host 0.0.0.0 --port 8900 > app.log 2>&1 & To kill sudo lsof -i :8900 sudo kill -9 858138 RUN THIS FILE ALWAYS WHENEVER YOUR DATASET UPGRADED python import_data.py ALSO CREATE FOLDER TO SAVE IMAGES mkdir uploads