feather.py

import pandas as pd
import os
from typing import Dict, List, Any, Optional
import math


class FeatherManager:
    
    def __init__(self, models_dir: str = "models"):
        self.models_dir = models_dir
        os.makedirs(models_dir, exist_ok=True)
    
    def save_mini_model(self, model_data: Dict[str, Any], model_id: int) -> str:
        filename = f"AgGPT_Expert_{model_id:04d}.feather"
        filepath = os.path.join(self.models_dir, filename)
        
        patterns = model_data.get('patterns', [])
        responses = model_data.get('responses', [])
        
        if not patterns or not responses:
            print(f"Warning: Model {model_id} has empty patterns or responses")
            patterns = patterns or ['hello']
            responses = responses or ['Hello!']
        
        # FIXED: Efficient data structure for AgGPT-19
        # Store only pattern-level data in main table
        df_data = {
            'patterns': [str(pattern) for pattern in patterns],
            'responses': [str(response) for response in responses],
            'weights': model_data.get('weights', [1.0] * len(patterns)),
            'response_templates': model_data.get('response_templates', [{}] * len(patterns)),
        }
        
        # Ensure all arrays are same length
        max_len = len(patterns)
        for key, value in df_data.items():
            if isinstance(value, list):
                while len(value) < max_len:
                    if key == 'weights':
                        value.append(1.0)
                    elif key == 'response_templates':
                        value.append({})
                    else:
                        value.append('')
        
        # Store model-level metadata separately (NO DUPLICATION!)
        model_metadata = {
            'model_id': model_id,
            'confidence': model_data.get('confidence', 0.5),
            'keywords': ' '.join(model_data.get('keywords', [])),
            'training_samples': model_data.get('training_samples', 0),
            'knowledge_base': str(model_data.get('knowledge_base', {})),
            'semantic_categories': str(model_data.get('semantic_categories', {})),
            'grammar_rules': str(model_data.get('grammar_rules', []))
        }
        
        # Create main patterns DataFrame
        df_patterns = pd.DataFrame(df_data)
        
        # Create single-row metadata DataFrame
        df_metadata = pd.DataFrame([model_metadata])
        
        # Save both as separate sheets in the same feather file using a different approach
        # Since feather doesn't support multiple sheets, we'll store metadata as the first row
        # and add a special marker
        
        # Create combined structure with metadata as first row
        metadata_row = {
            'patterns': f"__METADATA__{model_metadata['model_id']}",
            'responses': model_metadata['knowledge_base'],
            'weights': float(model_metadata['confidence']),
            'response_templates': f"{model_metadata['keywords']}|{model_metadata['training_samples']}|{model_metadata['semantic_categories']}|{model_metadata['grammar_rules']}"
        }
        
        # Combine metadata row with pattern data
        combined_data = [metadata_row]
        
        for i in range(len(patterns)):
            row = {
                'patterns': df_data['patterns'][i],
                'responses': df_data['responses'][i], 
                'weights': df_data['weights'][i],
                'response_templates': str(df_data['response_templates'][i])
            }
            combined_data.append(row)
        
        df_combined = pd.DataFrame(combined_data)
        df_combined.to_feather(filepath)
        
        print(f"Saved optimized mini-model: {filename} ({len(patterns)} patterns + metadata)")
        return filepath
    
    def load_mini_model(self, model_id: int) -> Optional[Dict[str, Any]]:
        filename = f"AgGPT_Expert_{model_id:04d}.feather"
        filepath = os.path.join(self.models_dir, filename)
        
        if not os.path.exists(filepath):
            return None
        
        try:
            df = pd.read_feather(filepath)
            
            # FIXED: Load optimized structure for AgGPT-19
            if len(df) == 0:
                return None
            
            # First row contains metadata (marked with __METADATA__)
            if df['patterns'].iloc[0].startswith('__METADATA__'):
                # Extract metadata from first row
                metadata_parts = df['response_templates'].iloc[0].split('|')
                
                model_data = {
                    'model_id': model_id,
                    'confidence': float(df['weights'].iloc[0]),
                    'keywords': metadata_parts[0].split() if len(metadata_parts) > 0 else [],
                    'training_samples': int(metadata_parts[1]) if len(metadata_parts) > 1 and metadata_parts[1].isdigit() else 0,
                }
                
                # Parse knowledge base from responses column
                try:
                    kb_str = df['responses'].iloc[0]
                    model_data['knowledge_base'] = eval(kb_str) if kb_str != '{}' else {}
                except:
                    model_data['knowledge_base'] = {}
                
                # Parse semantic categories
                try:
                    if len(metadata_parts) > 2:
                        model_data['semantic_categories'] = eval(metadata_parts[2]) if metadata_parts[2] != '{}' else {}
                    else:
                        model_data['semantic_categories'] = {}
                except:
                    model_data['semantic_categories'] = {}
                
                # Parse grammar rules
                try:
                    if len(metadata_parts) > 3:
                        model_data['grammar_rules'] = eval(metadata_parts[3]) if metadata_parts[3] != '[]' else []
                    else:
                        model_data['grammar_rules'] = []
                except:
                    model_data['grammar_rules'] = []
                
                # Extract pattern data (skip metadata row)
                pattern_df = df.iloc[1:].copy()
                
                model_data['patterns'] = [p for p in pattern_df['patterns'].tolist() if p]
                model_data['responses'] = [r for r in pattern_df['responses'].tolist() if r]
                model_data['weights'] = pattern_df['weights'].tolist()
                
                # Parse response templates
                response_templates = []
                for template_str in pattern_df['response_templates'].tolist():
                    try:
                        template = eval(template_str) if template_str not in ['{}', ''] else {}
                        response_templates.append(template)
                    except:
                        response_templates.append({})
                
                model_data['response_templates'] = response_templates
                
            else:
                # Fallback: Load old format (for backward compatibility)
                model_data = {
                    'patterns': [p for p in df['patterns'].tolist() if p],
                    'responses': [r for r in df['responses'].tolist() if r],
                    'weights': df['weights'].tolist(),
                    'confidence': df.get('confidence', pd.Series([0.5])).iloc[0] if 'confidence' in df.columns else 0.5,
                    'keywords': df.get('keywords', pd.Series([''])).iloc[0].split() if 'keywords' in df.columns else [],
                    'training_samples': df.get('training_samples', pd.Series([0])).iloc[0] if 'training_samples' in df.columns else 0,
                    'model_id': model_id,
                    'knowledge_base': {},
                    'semantic_categories': {},
                    'response_templates': [],
                    'grammar_rules': []
                }
            
            return model_data
            
        except Exception as e:
            print(f"Error loading optimized model {model_id}: {e}")
            return None
    
    def load_all_models(self) -> List[Dict[str, Any]]:
        models = []
        
        if not os.path.exists(self.models_dir):
            return models
        
        for filename in os.listdir(self.models_dir):
            if filename.startswith("AgGPT_Expert_") and filename.endswith(".feather"):
                try:
                    model_id = int(filename.split("_")[2].split(".")[0])
                    model = self.load_mini_model(model_id)
                    if model:
                        models.append(model)
                except (ValueError, IndexError):
                    print(f"Warning: Invalid model filename format: {filename}")
                    continue
        
        return models
    
    def get_model_count(self) -> int:
        if not os.path.exists(self.models_dir):
            return 0
        
        count = 0
        for filename in os.listdir(self.models_dir):
            if filename.startswith("AgGPT_Expert_") and filename.endswith(".feather"):
                count += 1
        
        return count
    
    def get_next_model_id(self) -> int:
        if not os.path.exists(self.models_dir):
            return 1
        
        max_id = 0
        for filename in os.listdir(self.models_dir):
            if filename.startswith("AgGPT_Expert_") and filename.endswith(".feather"):
                try:
                    model_id = int(filename.split("_")[2].split(".")[0])
                    max_id = max(max_id, model_id)
                except (ValueError, IndexError):
                    continue
        
        return max_id + 1
    
    def delete_model(self, model_id: int) -> bool:
        filename = f"AgGPT_Expert_{model_id:04d}.feather"
        filepath = os.path.join(self.models_dir, filename)
        
        if os.path.exists(filepath):
            try:
                os.remove(filepath)
                print(f"Deleted model: {filename}")
                return True
            except Exception as e:
                print(f"Error deleting model {model_id}: {e}")
                return False
        
        return False
    
    def clear_all_models(self) -> int:
        if not os.path.exists(self.models_dir):
            return 0
        
        deleted_count = 0
        for filename in os.listdir(self.models_dir):
            if filename.startswith("AgGPT_Expert_") and filename.endswith(".feather"):
                try:
                    os.remove(os.path.join(self.models_dir, filename))
                    deleted_count += 1
                except Exception as e:
                    print(f"Error deleting {filename}: {e}")
        
        print(f"Deleted {deleted_count} model files")
        return deleted_count


def similarity_score(text1: str, text2: str) -> float:
    """Enhanced semantic similarity calculation for AgGPT-19"""
    if not text1 or not text2:
        return 0.0
    
    # Normalize texts
    text1_clean = text1.lower().strip()
    text2_clean = text2.lower().strip()
    
    # Exact match bonus
    if text1_clean == text2_clean:
        return 1.0
    
    # Character-level similarity (for typos and variations)
    char_sim = _character_similarity(text1_clean, text2_clean)
    
    # Word-level analysis
    words1 = set(text1_clean.split())
    words2 = set(text2_clean.split())
    
    if not words1 or not words2:
        return char_sim * 0.3
    
    # Jaccard similarity (word overlap)
    intersection = len(words1.intersection(words2))
    union = len(words1.union(words2))
    jaccard = intersection / union if union > 0 else 0.0
    
    # Semantic word analysis
    semantic_sim = _semantic_word_similarity(words1, words2)
    
    # N-gram similarity
    ngram_sim = _ngram_similarity(text1_clean, text2_clean)
    
    # Length penalty for very different lengths
    len1, len2 = len(text1_clean.split()), len(text2_clean.split())
    length_penalty = 1.0 - min(abs(len1 - len2) / max(len1, len2, 1), 0.5)
    
    # Combine all similarity measures
    final_score = (
        jaccard * 0.4 +           # Word overlap
        semantic_sim * 0.3 +      # Semantic similarity
        ngram_sim * 0.2 +         # Character patterns
        char_sim * 0.1            # Character similarity
    ) * length_penalty
    
    return min(final_score, 1.0)


def _character_similarity(text1: str, text2: str) -> float:
    """Calculate character-level similarity using longest common subsequence"""
    if not text1 or not text2:
        return 0.0
    
    # Simple LCS implementation
    len1, len2 = len(text1), len(text2)
    dp = [[0] * (len2 + 1) for _ in range(len1 + 1)]
    
    for i in range(1, len1 + 1):
        for j in range(1, len2 + 1):
            if text1[i-1] == text2[j-1]:
                dp[i][j] = dp[i-1][j-1] + 1
            else:
                dp[i][j] = max(dp[i-1][j], dp[i][j-1])
    
    lcs_length = dp[len1][len2]
    return (2.0 * lcs_length) / (len1 + len2) if (len1 + len2) > 0 else 0.0


def _semantic_word_similarity(words1: set, words2: set) -> float:
    """Calculate semantic similarity between word sets"""
    if not words1 or not words2:
        return 0.0
    
    # Common semantic patterns
    synonyms = {
        'hello': {'hi', 'hey', 'greetings', 'good morning', 'good afternoon'},
        'thanks': {'thank you', 'appreciate', 'grateful'},
        'yes': {'yeah', 'yep', 'sure', 'absolutely', 'definitely'},
        'no': {'nope', 'negative', 'not really'},
        'good': {'great', 'excellent', 'wonderful', 'amazing', 'fantastic'},
        'bad': {'terrible', 'awful', 'horrible', 'poor'},
        'big': {'large', 'huge', 'enormous', 'massive'},
        'small': {'little', 'tiny', 'mini', 'minute'},
    }
    
    # Find semantic matches
    semantic_matches = 0
    total_comparisons = 0
    
    for word1 in words1:
        for word2 in words2:
            total_comparisons += 1
            
            # Direct match
            if word1 == word2:
                semantic_matches += 1
                continue
            
            # Check synonyms
            for key, synonym_set in synonyms.items():
                if word1 in synonym_set and word2 in synonym_set:
                    semantic_matches += 0.8
                    break
                elif (word1 == key and word2 in synonym_set) or (word2 == key and word1 in synonym_set):
                    semantic_matches += 0.9
                    break
            
            # Partial word matching (for variations)
            if len(word1) > 3 and len(word2) > 3:
                if word1 in word2 or word2 in word1:
                    semantic_matches += 0.6
                elif word1[:3] == word2[:3]:  # Same prefix
                    semantic_matches += 0.4
    
    return semantic_matches / total_comparisons if total_comparisons > 0 else 0.0


def _ngram_similarity(text1: str, text2: str, n: int = 3) -> float:
    """Calculate n-gram similarity for character patterns"""
    if len(text1) < n or len(text2) < n:
        return 0.0
    
    ngrams1 = set(text1[i:i+n] for i in range(len(text1) - n + 1))
    ngrams2 = set(text2[i:i+n] for i in range(len(text2) - n + 1))
    
    if not ngrams1 or not ngrams2:
        return 0.0
    
    intersection = len(ngrams1.intersection(ngrams2))
    union = len(ngrams1.union(ngrams2))
    
    return intersection / union if union > 0 else 0.0


def calculate_confidence_score(patterns: List[str], responses: List[str]) -> float:
    if not patterns or not responses or len(patterns) != len(responses):
        return 0.1
    
    base_confidence = min(0.9, len(patterns) / 10.0)
    
    return max(0.1, min(1.0, base_confidence))


if __name__ == "__main__":
    manager = FeatherManager()
    
    test_model = {
        'patterns': ['hello', 'hi', 'hey'],
        'responses': ['Hello! How can I help you?', 'Hi there!', 'Hey! What\'s up?'],
        'weights': [1.0, 0.9, 0.8],
        'confidence': 0.8,
        'grammar_rules': ['capitalize_first_word', 'end_with_punctuation'],
        'keywords': ['greeting', 'hello', 'hi'],
        'training_samples': 150
    }
    
    model_id = manager.get_next_model_id()
    manager.save_mini_model(test_model, model_id)
    loaded_model = manager.load_mini_model(model_id)
    
    print(f"Original model: {test_model}")
    print(f"Loaded model: {loaded_model}")
    print(f"Models count: {manager.get_model_count()}")