src/evaluation.py

"""
RAG Evaluation Module
Comprehensive evaluation system for RAG-based Q&A
"""
import pandas as pd
import numpy as np
from sentence_transformers import SentenceTransformer
from sklearn.metrics.pairwise import cosine_similarity
import nltk
import re
from typing import Dict, List, Optional
from datetime import datetime

# Download required NLTK data
try:
    nltk.download('punkt', quiet=True)
    nltk.download('stopwords', quiet=True)
except:
    pass


class ComprehensiveRAGEvaluator:
    """Comprehensive evaluation system for RAG-based car manual Q&A"""
    
    def __init__(self, rag_system, client):
        self.rag_system = rag_system
        self.client = client
        self.sentence_model = SentenceTransformer('all-MiniLM-L6-v2')
        self.evaluation_results = {}
    
    def evaluate_answer_quality(self, question: str, generated_answer: str, 
                                expected_answer: str, retrieved_contexts: List[str]) -> Dict:
        """
        Comprehensive answer quality evaluation
        
        Args:
            question: The question asked
            generated_answer: Answer generated by RAG system
            expected_answer: Expected correct answer
            retrieved_contexts: Contexts retrieved for the answer
            
        Returns:
            Dictionary of quality metrics
        """
        metrics = {}
        
        # 1. Semantic Similarity to Expected Answer
        gen_embedding = self.sentence_model.encode([generated_answer])
        exp_embedding = self.sentence_model.encode([expected_answer])
        metrics['semantic_similarity'] = cosine_similarity(gen_embedding, exp_embedding)[0][0]
        
        # 2. Answer Relevance to Question
        q_embedding = self.sentence_model.encode([question])
        a_embedding = self.sentence_model.encode([generated_answer])
        metrics['answer_relevance'] = cosine_similarity(q_embedding, a_embedding)[0][0]
        
        # 3. Faithfulness (grounding in retrieved context)
        metrics['faithfulness'] = self._calculate_faithfulness(generated_answer, retrieved_contexts)
        
        # 4. Completeness Assessment
        metrics['completeness'] = self._assess_completeness(question, generated_answer, expected_answer)
        
        # 5. Safety Appropriateness
        metrics['safety_appropriateness'] = self._check_safety_appropriateness(question, generated_answer)
        
        # 6. Technical Accuracy
        metrics['technical_accuracy'] = self._assess_technical_accuracy(generated_answer, retrieved_contexts)
        
        # 7. Clarity and Actionability
        metrics['clarity'] = self._assess_clarity(generated_answer)
        metrics['actionability'] = self._assess_actionability(question, generated_answer)
        
        return metrics
    
    def _calculate_faithfulness(self, answer: str, contexts: List[str]) -> float:
        """Calculate how well the answer is grounded in the retrieved contexts"""
        if not contexts:
            return 0.0
        
        answer_sentences = nltk.sent_tokenize(answer)
        supported_sentences = 0
        
        for sentence in answer_sentences:
            sentence_embedding = self.sentence_model.encode([sentence])
            max_similarity = 0
            
            for context in contexts:
                context_embedding = self.sentence_model.encode([context])
                similarity = cosine_similarity(sentence_embedding, context_embedding)[0][0]
                max_similarity = max(max_similarity, similarity)
            
            if max_similarity > 0.7:
                supported_sentences += 1
        
        return supported_sentences / len(answer_sentences) if answer_sentences else 0.0
    
    def _assess_completeness(self, question: str, generated_answer: str, expected_answer: str) -> float:
        """Assess if the generated answer covers all aspects of the expected answer"""
        expected_words = set(expected_answer.lower().split())
        generated_words = set(generated_answer.lower().split())
        
        stop_words = set(['the', 'a', 'an', 'and', 'or', 'but', 'in', 'on', 'at', 'to', 'for', 'of', 'with', 'by'])
        expected_words -= stop_words
        generated_words -= stop_words
        
        if not expected_words:
            return 1.0
        
        overlap = len(expected_words.intersection(generated_words))
        return overlap / len(expected_words)
    
    def _check_safety_appropriateness(self, question: str, answer: str) -> float:
        """Check if safety-critical information is handled appropriately"""
        safety_keywords = ['brake', 'airbag', 'emergency', 'warning', 'danger', 'caution', 'safety', 'speed', 'steering']
        
        question_lower = question.lower()
        answer_lower = answer.lower()
        
        is_safety_related = any(keyword in question_lower for keyword in safety_keywords)
        
        if not is_safety_related:
            return 1.0
        
        safety_indicators = ['warning', 'caution', 'important', 'ensure', 'never', 'always', 'must']
        has_safety_language = any(indicator in answer_lower for indicator in safety_indicators)
        
        return 1.0 if has_safety_language else 0.5
    
    def _assess_technical_accuracy(self, answer: str, contexts: List[str]) -> float:
        """Assess technical accuracy based on context alignment"""
        if not contexts:
            return 0.5
        
        answer_embedding = self.sentence_model.encode([answer])
        context_embeddings = self.sentence_model.encode(contexts)
        
        similarities = cosine_similarity(answer_embedding, context_embeddings)[0]
        return np.mean(similarities)
    
    def _assess_clarity(self, answer: str) -> float:
        """Assess clarity of the answer"""
        sentences = nltk.sent_tokenize(answer)
        
        if not sentences:
            return 0.0
        
        avg_sentence_length = np.mean([len(sentence.split()) for sentence in sentences])
        length_score = min(1.0, 15.0 / avg_sentence_length) if avg_sentence_length > 0 else 0.0
        
        structure_indicators = ['step', 'first', 'second', 'then', 'next', 'finally', '1.', '2.']
        has_structure = any(indicator in answer.lower() for indicator in structure_indicators)
        structure_score = 1.0 if has_structure else 0.7
        
        return (length_score + structure_score) / 2
    
    def _assess_actionability(self, question: str, answer: str) -> float:
        """Assess if the answer provides actionable information"""
        question_lower = question.lower()
        answer_lower = answer.lower()
        
        if 'how to' in question_lower or 'how do' in question_lower:
            action_indicators = ['press', 'turn', 'select', 'push', 'pull', 'set', 'adjust', 'follow', 'ensure']
            has_actions = any(indicator in answer_lower for indicator in action_indicators)
            return 1.0 if has_actions else 0.3
        
        return 0.8
    
    def generate_evaluation_report(self) -> str:
        """Generate comprehensive evaluation report"""
        if not self.evaluation_results:
            return "No evaluation results available. Run evaluation first."
        
        df = pd.DataFrame(self.evaluation_results)
        
        # Overall metrics
        overall_metrics = {
            'semantic_similarity': df['semantic_similarity'].mean(),
            'answer_relevance': df['answer_relevance'].mean(),
            'faithfulness': df['faithfulness'].mean(),
            'completeness': df['completeness'].mean(),
            'safety_appropriateness': df['safety_appropriateness'].mean(),
            'technical_accuracy': df['technical_accuracy'].mean(),
            'clarity': df['clarity'].mean(),
            'actionability': df['actionability'].mean()
        }
        
        # Performance by question type
        type_performance = df.groupby('question_type')[list(overall_metrics.keys())].mean()
        
        # Performance by difficulty
        difficulty_performance = df.groupby('difficulty')[list(overall_metrics.keys())].mean()
        
        # Generate report
        report = f"""
# RAG System Comprehensive Evaluation Report
Generated on: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}

## Overall Performance Metrics
{'-' * 40}
{'Metric':<25} {'Score':<10} {'Interpretation':<30}
{'-' * 40}
{'Semantic Similarity':<25} {overall_metrics['semantic_similarity']:.3f}      {'Answer matches expected content'}
{'Answer Relevance':<25} {overall_metrics['answer_relevance']:.3f}      {'Answer addresses the question'}
{'Faithfulness':<25} {overall_metrics['faithfulness']:.3f}      {'Answer is grounded in context'}
{'Completeness':<25} {overall_metrics['completeness']:.3f}      {'Answer covers all aspects'}
{'Safety Appropriateness':<25} {overall_metrics['safety_appropriateness']:.3f}      {'Safety info handled properly'}
{'Technical Accuracy':<25} {overall_metrics['technical_accuracy']:.3f}      {'Technically correct information'}
{'Clarity':<25} {overall_metrics['clarity']:.3f}      {'Clear and understandable'}
{'Actionability':<25} {overall_metrics['actionability']:.3f}      {'Provides actionable guidance'}
{'-' * 40}

## Performance by Question Type
{type_performance.round(3)}

## Performance by Difficulty Level
{difficulty_performance.round(3)}
"""
        
        return report