backend/validate_ai_logic.py

import sys
import os
import pandas as pd
import statistics

# Add current directory to path so we can import app
sys.path.append(os.getcwd())

from app.services.data_service import data_service

def run_validation():
    print("Loading data...")
    data_service.load_data()
    
    if data_service.master_df is None:
        print("Error: Could not load data.")
        return

    # Get unique articles from master_df
    # Column might be 'Article No' or 'Article' - check data_service.load_data
    # Usually it's mapped. Let's start with 'Article' if mapped, or 'Material'
    # In data_service.py, keys are likely from CSV.
    # Let's verify column name. SAFE bet is to iterate what get_article_predictions expects.
    # It takes article_id.
    # Let's try to get unique values from the dataframe.
    
    df = data_service.master_df
    # Identify article column
    article_col = 'Article' if 'Article' in df.columns else 'Material' 
    # If not found, try to list columns
    if article_col not in df.columns:
        print(f"Columns: {df.columns}")
        return

    all_articles = df[article_col].dropna().unique()
    print(f"Found {len(all_articles)} unique articles. Running AI validation...")
    
    results = []
    
    # Run for all
    count = 0
    for article in all_articles:
        count += 1
        if count % 100 == 0:
            print(f"Processed {count}/{len(all_articles)}...")
            
        try:
            pred = data_service.get_article_predictions(article)
            ai = pred['ai_prediction']
            rec = ai['recommendation']
            norm = ai['norm_analysis']
            stats = ai['historical_analysis']
            yield_stats = ai['yield_stats']
            
            results.append({
                'Article': article,
                'OrderCount': ai['historical_orders'],
                'SuccessRate': stats['success_rate_pct'],
                'NormPct': norm['base_norm_pct'],
                'RecPct': rec['suggested_reservation_pct'],
                'Adjustment': rec['ai_adjustment_pct'],
                'YieldAvg': yield_stats['avg'],
                'YieldStd': yield_stats['std_dev']
            })
        except Exception as e:
            print(f"Error processing {article}: {e}")
            pass

    print(f"Collected results for {len(results)} articles.")
    if not results:
        print("No results collected! Exiting.")
        return

    res_df = pd.DataFrame(results)
    res_df['Savings'] = res_df['NormPct'] - res_df['RecPct']
    
    # Save to CSV for inspection
    res_df.to_csv('validation_results.csv', index=False)
    
    print("\n" + "="*40)
    print("       VALIDATION SUMMARY")
    print("="*40)
    print(f"Total Articles Analyzed: {len(res_df)}")
    print(f"Articles with Savings (> 0.5%): {len(res_df[res_df['Savings'] > 0.5])}")
    print(f"Articles with More Buffer (< -0.5%): {len(res_df[res_df['Savings'] < -0.5])}")
    print(f"Average Savings across plant: {res_df['Savings'].mean():.2f}%")
    
    print("\n--- TOP 5 SAVINGS OPPORTUNITIES (Less Waste) ---")
    print(res_df[res_df['OrderCount'] > 5].sort_values('Savings', ascending=False).head(5)[['Article', 'OrderCount', 'NormPct', 'RecPct', 'Savings', 'SuccessRate']])
    
    print("\n--- TOP 5 RISK MITIGATION (More Safety) ---")
    print(res_df[res_df['OrderCount'] > 5].sort_values('Savings', ascending=True).head(5)[['Article', 'OrderCount', 'NormPct', 'RecPct', 'Savings', 'SuccessRate']])
    
    # Anomalies
    neg_recs = res_df[res_df['RecPct'] < 0]
    if not neg_recs.empty:
        print(f"\n[CRITICAL] Found {len(neg_recs)} articles with NEGATIVE recommendation!")
        print(neg_recs[['Article', 'RecPct']])
    
    high_recs = res_df[res_df['RecPct'] > 15]
    if not high_recs.empty:
        print(f"\n[WARNING] Found {len(high_recs)} articles with > 15% recommendation!")
        print(high_recs[['Article', 'RecPct', 'OrderCount']])

    # Check specifically for the 'Partial Order' edge cases (high yields but high failure rate if not handled)
    # We can't easily identify them here without looking at raw orders, but we can see if Rec % is reasonable.

if __name__ == "__main__":
    run_validation()