Fix structure detection - require keywords for opening/closing rules, prevent false positive from blind confidence trust

app/services/structure.py

@@ -94,29 +94,13 @@ class StructureService:
         
         n = len(predictions)
         
-        # Rule 1: 2 kalimat pertama cenderung opening
-        for i in range(min(2, n)):
-            if predictions[i]['confidence'] > 0.5:
-                probs_opening = predictions[i].get('confidence', 0)
-                if probs_opening > 0.8:
-                    predictions[i]['predicted_label'] = 'opening'
-        
-        # Rule 2: 2 kalimat terakhir cenderung closing
-        for i in range(max(0, n-2), n):
-            if predictions[i]['confidence'] > 0.5:
-                probs_closing = predictions[i].get('confidence', 0)
-                if probs_closing > 0.8:
-                    predictions[i]['predicted_label'] = 'closing'
-        
-        # Rule 3: Detect keywords dengan exact matching
-        # CLOSING: Hanya kalimat dengan kata kunci penutup yang JELAS
+        # Define keywords first (will be used in rules)
         closing_keywords = [
             'demikian', 'terima kasih', 'terimakasih', 'sekian', 'akhir kata',
             'wassalamualaikum', 'wassalam', 'waalaikumsalam',
             'sampai jumpa', 'sampai bertemu', 'salam penutup'
         ]
         
-        # OPENING: Kata kunci pembuka
         opening_keywords = [
             'selamat pagi', 'selamat siang', 'selamat sore', 'selamat malam',
             'assalamualaikum', 'assalamu alaikum',
@@ -124,6 +108,23 @@ class StructureService:
             'perkenalkan', 'yang terhormat'
         ]
         
+        # Rule 1: 2 kalimat pertama cenderung opening (HANYA jika ada opening keyword)
+        for i in range(min(2, n)):
+            text_lower = predictions[i]['text'].lower()
+            has_opening_kw = any(kw in text_lower for kw in opening_keywords)
+            
+            if has_opening_kw and predictions[i]['confidence'] > 0.5:
+                predictions[i]['predicted_label'] = 'opening'
+        
+        # Rule 2: 2 kalimat terakhir cenderung closing (HANYA jika ada closing keyword)
+        for i in range(max(0, n-2), n):
+            text_lower = predictions[i]['text'].lower()
+            has_closing_kw = any(kw in text_lower for kw in closing_keywords)
+            
+            if has_closing_kw and predictions[i]['confidence'] > 0.5:
+                predictions[i]['predicted_label'] = 'closing'
+        
+        # Rule 3: Keyword detection untuk semua kalimat (override model prediction)
         for pred in predictions:
             text_lower = pred['text'].lower()