Update app.py

app.py

@@ -275,7 +275,7 @@ async def download_voice_file(media_url: str, filename: str) -> str:
         return None
 
 async def transcribe_voice_with_openai(file_path: str) -> str:
-    """Transcribe voice file using OpenAI Whisper with comprehensive veterinary domain system prompt"""
+    """Transcribe voice file using OpenAI Whisper with intelligent English/Urdu focus"""
     try:
         # Check if file exists and has content
         if not os.path.exists(file_path):
@@ -289,19 +289,11 @@ async def transcribe_voice_with_openai(file_path: str) -> str:
         
         logger.info(f"[Transcribe] Transcribing file: {file_path} (size: {file_size} bytes)")
         
-        # Comprehensive system prompt for veterinary WhatsApp assistant - ENGLISH AND URDU ONLY
+        # Intelligent English/Urdu focused system prompt
         system_prompt = """
-You are transcribing voice messages for Apex Biotical Veterinary WhatsApp Assistant. This is a professional veterinary products chatbot.
+You are transcribing voice messages for Apex Biotical Veterinary WhatsApp Assistant.
 
-CRITICAL: TRANSCRIBE ONLY ENGLISH OR URDU SPEECH - REJECT ALL OTHER LANGUAGES
-
-IMPORTANT RULES:
-1. ONLY transcribe English or Urdu speech
-2. If you hear any other language, transcribe as "unclear audio"
-3. If you hear unclear audio, transcribe as "unclear audio"
-4. Never transcribe gibberish or random characters
-5. Keep transcriptions simple and clean
-6. Reject non-English/Urdu languages completely
+FOCUS: The user will speak in English, Urdu, or a mix of both languages. Be intelligent and natural in understanding their speech.
 
 PRODUCT NAMES (exact spelling required):
 - Hydropex, Respira Aid Plus, Heposel, Bromacid, Hexatox
@@ -310,22 +302,25 @@ PRODUCT NAMES (exact spelling required):
 - Apvita Plus, B-G Aspro-C, EC-Immune, Liverpex, Symodex
 - Respira Aid, Adek Gold, Immuno DX
 
+ENGLISH NUMBERS: one->1, two->2, three->3, four->4, five->5, six->6, seven->7, eight->8, nine->9, ten->10
+
+URDU NUMBERS: aik->1, ek->1, do->2, teen->3, char->4, panch->5, cheh->6, saat->7, aath->8, nau->9, das->10
+
 MENU COMMANDS:
-- Numbers: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
-- Navigation: main, menu, back, home, start
-- Options: option, number, choice, select
+- English: main, menu, back, home, start, option, number, search, browse, download, catalog, contact, availability
+- Urdu: main menu, option, number, search, browse, download, catalog, contact, availability
 
 GREETINGS:
 - English: hi, hello, hey, good morning, good afternoon, good evening
 - Urdu: salam, assalamu alaikum, adaab, namaste, khuda hafiz
 
 TRANSCRIPTION RULES:
-1. Transcribe exactly what you hear in English or Urdu ONLY
-2. Convert numbers to digits (one->1, two->2, etc.)
-3. Preserve product names exactly
-4. If unclear or non-English/Urdu, transcribe as "unclear audio"
-5. Keep it simple and clean
-6. Reject all other languages
+1. Intelligently transcribe English and Urdu speech
+2. Handle mixed English-Urdu speech naturally
+3. Convert numbers to digits
+4. Preserve product names exactly
+5. Only return "unclear audio" if the voice is genuinely unclear or inaudible
+6. Be natural and conversational in understanding
 
 EXAMPLES:
 - "hydropex" -> "hydropex"
@@ -334,11 +329,12 @@ EXAMPLES:
 - "main menu" -> "main"
 - "salam" -> "salam"
 - "search products" -> "search products"
-- Non-English/Urdu speech -> "unclear audio"
-- Unclear audio -> "unclear audio"
+- "how many products" -> "how many products"
+- "kitne products hain" -> "kitne products hain"
+- Genuinely unclear audio -> "unclear audio"
 """
         
-        # First attempt with comprehensive system prompt
+        # First attempt with intelligent English/Urdu focus
         with open(file_path, 'rb') as audio_file:
             transcript = openai.Audio.transcribe(
                 model="whisper-1",
@@ -357,6 +353,8 @@ EXAMPLES:
             urdu_system_prompt = """
 You are transcribing Urdu voice messages for Apex Biotical Veterinary WhatsApp Assistant.
 
+FOCUS: The user will speak in Urdu, English, or a mix of both. Be intelligent and natural.
+
 PRODUCT NAMES (Urdu/English):
 - ہائیڈروپیکس (Hydropex)
 - ریسپیرا ایڈ پلس (Respira Aid Plus)
@@ -369,12 +367,11 @@ PRODUCT NAMES (Urdu/English):
 - فائٹو سال (PHYTO-SAL)
 - مائیکوپیکس سپر (Mycopex Super)
 
-URDU NUMBERS:
+URDU NUMBERS (convert to digits):
 - ایک (1), دو (2), تین (3), چار (4), پانچ (5)
 - چھ (6), سات (7), آٹھ (8), نو (9), دس (10)
 - گیارہ (11), بارہ (12), تیرہ (13), چودہ (14), پندرہ (15)
 - سولہ (16), سترہ (17), اٹھارہ (18), انیس (19), بیس (20)
-- اکیس (21), بائیس (22), تئیس (23)
 
 URDU GREETINGS:
 - سلام (salam), السلام علیکم (assalamu alaikum)
@@ -386,11 +383,12 @@ URDU MENU COMMANDS:
 - کیٹلاگ (catalog), رابطہ (contact), دستیابی (availability)
 
 TRANSCRIPTION RULES:
-1. Transcribe Urdu words in Urdu script
-2. Convert Urdu numbers to digits
-3. Handle mixed Urdu-English speech
+1. Intelligently transcribe Urdu and English speech
+2. Handle mixed language naturally
+3. Convert Urdu numbers to digits
 4. Preserve product names exactly
-5. Convert menu selections to numbers
+5. Only return "unclear audio" if voice is genuinely unclear
+6. Be natural and conversational
 """
             
             with open(file_path, 'rb') as audio_file:
@@ -403,59 +401,16 @@ TRANSCRIPTION RULES:
                 
             transcribed_text = transcript.text.strip()
             logger.info(f"[Transcribe] Second attempt transcribed (Urdu): '{transcribed_text}'")
-            
-            # Third attempt with mixed language prompt if still failing
-            if not transcribed_text or len(transcribed_text.strip()) < 2:
-                logger.warning(f"[Transcribe] Second attempt failed, trying with mixed language prompt")
-                
-                mixed_system_prompt = """
-You are transcribing voice messages for a veterinary products WhatsApp assistant. The user may speak in English, Urdu, or a mix of both languages.
-
-PRODUCT NAMES (exact spelling required):
-Hydropex, Respira Aid Plus, Heposel, Bromacid, Hexatox, APMA Fort, Para C.E, Tribiotic, PHYTO-SAL, Mycopex Super, Eflin KT-20, Salcozine ST-30, Oftilex UA-10, Biscomin 10, Apvita Plus, B-G Aspro-C, EC-Immune, Liverpex, Symodex, Respira Aid, Adek Gold, Immuno DX
-
-NUMBERS (convert to digits):
-English: one->1, two->2, three->3, etc.
-Urdu: aik->1, ek->1, do->2, teen->3, etc.
-
-MENU COMMANDS:
-main, menu, back, home, start, option, number, search, browse, download, catalog, contact, availability
-
-GREETINGS:
-hi, hello, salam, assalamu alaikum, adaab, namaste
-
-TRANSCRIPTION RULES:
-1. Transcribe exactly what you hear
-2. Convert numbers to digits
-3. Preserve product names exactly
-4. Handle both languages
-5. Convert menu selections to numbers
-"""
-                
-                with open(file_path, 'rb') as audio_file:
-                    transcript = openai.Audio.transcribe(
-                        model="whisper-1",
-                        file=audio_file,
-                        prompt=mixed_system_prompt
-                    )
-                    
-                transcribed_text = transcript.text.strip()
-                logger.info(f"[Transcribe] Third attempt (mixed) transcribed: '{transcribed_text}'")
         
-        # Final check for empty transcription or unclear audio
+        # Final check for genuinely unclear audio
         if not transcribed_text or len(transcribed_text.strip()) < 2:
             logger.warning(f"[Transcribe] Very short or empty transcription: '{transcribed_text}'")
             return "unclear audio"
         
-        # Check for gibberish or mixed characters
-        if len(transcribed_text) > 10 and not re.search(r'[a-zA-Z\u0600-\u06FF]', transcribed_text):
-            logger.warning(f"[Transcribe] Gibberish detected: '{transcribed_text}'")
-            return "unclear audio"
-        
-        # Check for too many special characters
+        # Check for too many special characters (indicates unclear audio)
         special_char_ratio = len(re.findall(r'[^\w\s]', transcribed_text)) / len(transcribed_text)
-        if special_char_ratio > 0.3:
-            logger.warning(f"[Transcribe] Too many special characters: '{transcribed_text}'")
+        if special_char_ratio > 0.5:  # More than 50% special characters
+            logger.warning(f"[Transcribe] Too many special characters, unclear audio: '{transcribed_text}'")
             return "unclear audio"
         
         return transcribed_text