Update process_interview.py

process_interview.py

@@ -69,7 +69,6 @@ def download_audio_from_url(url: str) -> str:
 
 
 
-# Initialize services
 def initialize_services():
     # Pinecone
     pc = Pinecone(api_key=PINECONE_KEY)
@@ -83,12 +82,17 @@ def initialize_services():
         )
     index = pc.Index(index_name)
     
+    # حذف أي بيانات قديمة (اختياري)
+    try:
+        index.delete(delete_all=True)
+    except Exception as e:
+        logger.warning(f"Could not clear index: {str(e)}")
+    
     # Gemini
     genai.configure(api_key=GEMINI_API_KEY)
     gemini_model = genai.GenerativeModel('gemini-1.5-flash')
     
     return index, gemini_model
-
 index, gemini_model = initialize_services()
 
 # Device setup
@@ -228,13 +232,14 @@ def process_utterance(utterance, full_audio, wav_file):
         )
         
         # Identify speaker
-        if query_result['matches'] and query_result['matches'][0]['score'] > 0.7:
+        if query_result['matches'] and query_result['matches'][0]['score'] > 0.5:  # تخفيض العتبة
             speaker_id = query_result['matches'][0]['id']
             speaker_name = query_result['matches'][0]['metadata']['speaker_name']
         else:
             speaker_id = f"unknown_{uuid.uuid4().hex[:6]}"
             speaker_name = f"Speaker_{speaker_id[-4:]}"
-            index.upsert([(speaker_id, embedding.tolist(), {"speaker_name": speaker_name})])
+            # إضافة المتحدث الجديد إلى الفهرس
+            index.upsert([(speaker_id, embedding.cpu().numpy().tolist(), {"speaker_name": speaker_name})])
         
         # Cleanup
         os.remove(temp_path)
@@ -275,7 +280,7 @@ def identify_speakers(transcript: Dict, wav_file: str) -> List[Dict]:
 # Role classification
 def train_role_classifier(utterances: List[Dict]):
     try:
-        # Prepare data
+        # تحليل المحتوى للتمييز بين الأسئلة (المحاور) والإجابات (المتحدث)
         texts = [u['text'] for u in utterances]
         vectorizer = TfidfVectorizer(max_features=500, ngram_range=(1, 2))
         X_text = vectorizer.fit_transform(texts)
@@ -303,16 +308,19 @@ def train_role_classifier(utterances: List[Dict]):
             
             # Linguistic features
             doc = nlp(utterance['text'])
+            is_question = int(utterance['text'].endswith('?'))
+            question_words = len(re.findall(r'\b(why|how|what|when|where|who|which)\b', utterance['text'].lower()))
             feat.extend([
-                int(utterance['text'].endswith('?')),
-                len(re.findall(r'\b(why|how|what|when|where|who|which)\b', utterance['text'].lower())),
+                is_question,
+                question_words,
                 len(utterance['text'].split()),
                 sum(1 for token in doc if token.pos_ == 'VERB'),
                 sum(1 for token in doc if token.pos_ == 'NOUN')
             ])
             
             features.append(feat)
-            labels.append(0 if i % 2 == 0 else 1)  # Temporary labeling
+            # التصنيف بناءً على كون النص سؤالاً (محاور) أو لا (متحدث)
+            labels.append(0 if is_question or question_words > 0 else 1)
         
         # Train classifier
         scaler = StandardScaler()
@@ -336,7 +344,6 @@ def train_role_classifier(utterances: List[Dict]):
         logger.error(f"Classifier training failed: {str(e)}")
         raise
 
-
 def classify_roles(utterances: List[Dict], clf, vectorizer, scaler):
     try:
         # Prepare features for classification