Spaces:
Runtime error
Runtime error
| # models/trust_score.py | |
| from .model_loader import load_model | |
| from .logging_config import logger | |
| def generate_trust_score(text, image_analysis, pdf_analysis): | |
| try: | |
| try: | |
| classifier = load_model("zero-shot-classification", "typeform/mobilebert-uncased-mnli") | |
| except Exception as e: | |
| logger.error(f"Error loading model in trust score: {str(e)}") | |
| return 20, f"Model loading error: {str(e)}" | |
| aspects = [ | |
| "complete information provided", | |
| "verified location", | |
| "consistent data", | |
| "authentic documents", | |
| "authentic images", | |
| "reasonable pricing", | |
| "verified ownership", | |
| "proper documentation" | |
| ] | |
| try: | |
| result = classifier(str(text)[:1000], aspects, multi_label=True) | |
| except Exception as e: | |
| logger.error(f"Error in trust score model inference: {str(e)}") | |
| return 20, f"Model inference error: {str(e)}" | |
| # Much stricter weights with higher emphasis on critical aspects | |
| weights = { | |
| "complete information provided": 0.25, | |
| "verified location": 0.20, | |
| "consistent data": 0.15, | |
| "authentic documents": 0.15, | |
| "authentic images": 0.10, | |
| "reasonable pricing": 0.05, | |
| "verified ownership": 0.05, | |
| "proper documentation": 0.05 | |
| } | |
| score = 0 | |
| reasoning_parts = [] | |
| # Much stricter scoring for each aspect | |
| for label, confidence in zip(result['labels'], result['scores']): | |
| adjusted_confidence = confidence | |
| # Stricter document verification | |
| if label == "authentic documents": | |
| if not pdf_analysis or len(pdf_analysis) == 0: | |
| adjusted_confidence = 0.0 | |
| else: | |
| doc_scores = [p.get('verification_score', 0) for p in pdf_analysis] | |
| adjusted_confidence = sum(doc_scores) / max(1, len(doc_scores)) | |
| # Heavily penalize if any document has low verification score | |
| if any(score < 0.7 for score in doc_scores): | |
| adjusted_confidence *= 0.4 | |
| # Additional penalty for missing documents | |
| if len(doc_scores) < 2: | |
| adjusted_confidence *= 0.5 | |
| # Stricter image verification | |
| elif label == "authentic images": | |
| if not image_analysis or len(image_analysis) == 0: | |
| adjusted_confidence = 0.0 | |
| else: | |
| img_scores = [i.get('authenticity_score', 0) for i in image_analysis] | |
| adjusted_confidence = sum(img_scores) / max(1, len(img_scores)) | |
| # Heavily penalize if any image has low authenticity score | |
| if any(score < 0.8 for score in img_scores): | |
| adjusted_confidence *= 0.4 | |
| # Additional penalty for AI-generated images | |
| if any(i.get('is_ai_generated', False) for i in image_analysis): | |
| adjusted_confidence *= 0.5 | |
| # Additional penalty for non-property related images | |
| if any(not i.get('is_property_related', False) for i in image_analysis): | |
| adjusted_confidence *= 0.6 | |
| # Stricter consistency check | |
| elif label == "consistent data": | |
| # Check for inconsistencies in the data | |
| if "inconsistent" in text.lower() or "suspicious" in text.lower(): | |
| adjusted_confidence *= 0.3 | |
| # Check for impossible values | |
| if "impossible" in text.lower() or "invalid" in text.lower(): | |
| adjusted_confidence *= 0.2 | |
| # Check for missing critical information | |
| if "missing" in text.lower() or "not provided" in text.lower(): | |
| adjusted_confidence *= 0.4 | |
| # Stricter completeness check | |
| elif label == "complete information provided": | |
| # Check for missing critical information | |
| if len(text) < 300 or "not provided" in text.lower() or "missing" in text.lower(): | |
| adjusted_confidence *= 0.4 | |
| # Check for vague or generic descriptions | |
| if "generic" in text.lower() or "vague" in text.lower(): | |
| adjusted_confidence *= 0.5 | |
| # Check for suspiciously short descriptions | |
| if len(text) < 150: | |
| adjusted_confidence *= 0.3 | |
| score += adjusted_confidence * weights.get(label, 0.1) | |
| reasoning_parts.append(f"{label} ({adjusted_confidence:.0%})") | |
| # Apply additional penalties for suspicious patterns | |
| if "suspicious" in text.lower() or "fraudulent" in text.lower(): | |
| score *= 0.5 | |
| # Apply penalties for suspiciously low values | |
| if "suspiciously low" in text.lower() or "unusually small" in text.lower(): | |
| score *= 0.6 | |
| # Apply penalties for inconsistencies | |
| if "inconsistent" in text.lower() or "mismatch" in text.lower(): | |
| score *= 0.6 | |
| # Apply penalties for missing critical information | |
| if "missing critical" in text.lower() or "incomplete" in text.lower(): | |
| score *= 0.7 | |
| # Ensure score is between 0 and 100 | |
| score = min(100, max(0, int(score * 100))) | |
| reasoning = f"Based on: {', '.join(reasoning_parts)}" | |
| return score, reasoning | |
| except Exception as e: | |
| logger.error(f"Error generating trust score: {str(e)}") | |
| return 20, "Could not assess trust." | |