Update app.py

app.py

@@ -886,7 +886,7 @@ def detect_explicit_abuse(text):
             return True
     
     return False
-@spaces.GPU
+
 @spaces.GPU
 def analyze_single_message(text, thresholds):
     """Analyze a single message for abuse patterns with boundary assessment"""
@@ -898,52 +898,51 @@ def analyze_single_message(text, thresholds):
             logger.debug("Empty text, returning zeros")
             return 0.0, [], [], {"label": "none"}, 1, 0.0, None, {'assessment': 'neutral', 'confidence': 0.5}
             
-    # BOUNDARY HEALTH CHECK - Add this new section
-    logger.debug("\n🛡️ BOUNDARY HEALTH ANALYSIS")
-    logger.debug("-" * 40)
-    boundary_class, boundary_confidence = predict_boundary_health(text)  # Unpack the tuple
-    boundary_assessment = get_boundary_assessment(text, boundary_class, boundary_confidence)
-    logger.debug(f"Boundary Class: {boundary_class} ({['Respected', 'Violated', 'Dismissed', 'Manipulative'][boundary_class]})")
-    logger.debug(f"Boundary Confidence: {boundary_confidence:.3f}")
-    logger.debug(f"Boundary Assessment: {boundary_assessment['label']}")
-    
-    # Get sentiment EARLY - BEFORE any early returns
-    sent_inputs = sentiment_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
-    sent_inputs = {k: v.to(device) for k, v in sent_inputs.items()}
-    with torch.no_grad():
-        sent_logits = sentiment_model(**sent_inputs).logits[0]
-        sent_probs = torch.softmax(sent_logits, dim=-1).cpu().numpy()
-    
-    # Add detailed logging
-    logger.debug("\n🎭 SENTIMENT ANALYSIS DETAILS")
-    logger.debug(f"Raw logits: {sent_logits}")
-    logger.debug(f"Probabilities: supportive={sent_probs[0]:.3f}, undermining={sent_probs[1]:.3f}")
-    
-    # Make sure we're using the correct index mapping
-    sentiment = SENTIMENT_LABELS[int(np.argmax(sent_probs))]
-    logger.debug(f"Selected sentiment: {sentiment}")
-    
-    # UPDATE THE OVERRIDE CONDITION:
-    # Now we need to check if boundary_class == 0 (Respected) instead of checking a probability
-    if (boundary_class != 0 and  # Not "Respected"
-        boundary_confidence < 0.7 and 
-        sentiment == "supportive" and 
-        len(text.split()) > 50 and
-        any(phrase in text.lower() for phrase in [
-            "i need you to", "i want to understand", "this isn't about", 
-            "about accuracy", "willing to do something different"
-        ])):
-    
-        logger.debug("🔄 Boundary assessment override: Sophisticated healthy boundary detected")
-        boundary_assessment = {
-            'assessment': 'healthy',
-            'label': 'Healthy Boundary (Sophisticated)',
-            'confidence': 0.85,
-            'description': 'Complex but healthy boundary-setting communication',
-            'recommendations': ['Continue this thoughtful, direct approach']
-        }
+        # BOUNDARY HEALTH CHECK - Add this new section
+        logger.debug("\n🛡️ BOUNDARY HEALTH ANALYSIS")
+        logger.debug("-" * 40)
+        boundary_class, boundary_confidence = predict_boundary_health(text)  # Unpack the tuple
+        boundary_assessment = get_boundary_assessment(text, boundary_class, boundary_confidence)
+        logger.debug(f"Boundary Class: {boundary_class} ({['Respected', 'Violated', 'Dismissed', 'Manipulative'][boundary_class]})")
+        logger.debug(f"Boundary Confidence: {boundary_confidence:.3f}")
+        logger.debug(f"Boundary Assessment: {boundary_assessment['label']}")
+        
+        # Get sentiment EARLY - BEFORE any early returns
+        sent_inputs = sentiment_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
+        sent_inputs = {k: v.to(device) for k, v in sent_inputs.items()}
+        with torch.no_grad():
+            sent_logits = sentiment_model(**sent_inputs).logits[0]
+            sent_probs = torch.softmax(sent_logits, dim=-1).cpu().numpy()
+        
+        # Add detailed logging
+        logger.debug("\n🎭 SENTIMENT ANALYSIS DETAILS")
+        logger.debug(f"Raw logits: {sent_logits}")
+        logger.debug(f"Probabilities: supportive={sent_probs[0]:.3f}, undermining={sent_probs[1]:.3f}")
+        
+        # Make sure we're using the correct index mapping
+        sentiment = SENTIMENT_LABELS[int(np.argmax(sent_probs))]
+        logger.debug(f"Selected sentiment: {sentiment}")
+        
+        # UPDATE THE OVERRIDE CONDITION:
+        # Now we need to check if boundary_class == 0 (Respected) instead of checking a probability
+        if (boundary_class != 0 and  # Not "Respected"
+            boundary_confidence < 0.7 and 
+            sentiment == "supportive" and 
+            len(text.split()) > 50 and
+            any(phrase in text.lower() for phrase in [
+                "i need you to", "i want to understand", "this isn't about", 
+                "about accuracy", "willing to do something different"
+            ])):
+        
+            logger.debug("🔄 Boundary assessment override: Sophisticated healthy boundary detected")
+            boundary_assessment = {
+                'assessment': 'healthy',
+                'label': 'Healthy Boundary (Sophisticated)',
+                'confidence': 0.85,
+                'description': 'Complex but healthy boundary-setting communication',
+                'recommendations': ['Continue this thoughtful, direct approach']
+            }
 
-        
         # EARLY SUPPORTIVE MESSAGE CHECK
         innocent_indicators = [
             'broken', 'not working', 'cracked', 'glass', 'screen', 'phone',
@@ -952,10 +951,11 @@ def analyze_single_message(text, thresholds):
         ]
             
         # Enhanced early return check - now includes boundary health
+        # NOTE: Changed healthy_prob to boundary_class == 0 (Respected)
         if (any(indicator in text.lower() for indicator in innocent_indicators) and 
             len(text.split()) < 20 and 
             not any(threat in text.lower() for threat in ['kill', 'hurt', 'destroy', 'hate']) and
-            healthy_prob > 0.7):  # Added boundary health check
+            boundary_class == 0):  # Updated boundary health check
         
             # If sentiment is strongly supportive AND boundary health is good, return early
             if sent_probs[0] > 0.8:  # 80% supportive
@@ -1025,11 +1025,12 @@ def analyze_single_message(text, thresholds):
             abuse_score = max(abuse_score, 70.0)
             
         # Apply boundary health modifier to abuse score
-        if healthy_prob > 0.8 and not explicit_abuse:
-            # Very healthy boundaries - cap abuse score much lower
+        # NOTE: Updated to use boundary_class instead of healthy_prob
+        if boundary_class == 0 and boundary_confidence > 0.8 and not explicit_abuse:
+            # Very healthy boundaries (Respected) - cap abuse score much lower
             abuse_score = min(abuse_score, 20.0)
             logger.debug(f"Capped abuse score to {abuse_score} due to very healthy boundaries")
-        elif healthy_prob > 0.6 and sentiment == "supportive":
+        elif boundary_class == 0 and boundary_confidence > 0.6 and sentiment == "supportive":
             # Moderately healthy boundaries with supportive sentiment
             abuse_score = min(abuse_score, 35.0)
             logger.debug(f"Capped abuse score to {abuse_score} due to healthy boundaries")
@@ -1058,8 +1059,9 @@ def analyze_single_message(text, thresholds):
         log_emotional_tone_usage(tone_tag, threshold_labels)
         
         # Check for the specific combination (final safety check)
+        # NOTE: Updated to use boundary_class instead of healthy_prob
         highest_pattern = max(matched_scores, key=lambda x: x[1])[0] if matched_scores else None
-        if sentiment == "supportive" and tone_tag == "neutral" and highest_pattern == "obscure language" and healthy_prob > 0.6:
+        if sentiment == "supportive" and tone_tag == "neutral" and highest_pattern == "obscure language" and boundary_class == 0:
             logger.debug("Message classified as likely non-abusive (supportive, neutral, healthy boundaries). Returning low risk.")
             return 0.0, [], [], {"label": "supportive"}, 1, 0.0, "neutral", boundary_assessment
 
@@ -1075,6 +1077,7 @@ def analyze_single_message(text, thresholds):
         logger.error(f"Error in analyze_single_message: {e}")
         logger.error(f"Traceback: {traceback.format_exc()}")
         return 0.0, [], [], {"label": "error"}, 1, 0.0, None, {'assessment': 'error', 'confidence': 0.0}
+
 def generate_abuse_score_chart(dates, scores, patterns):
     """Generate a timeline chart of abuse scores"""
     try: