tashkeel fixer under test

self_hosted_env/voice_agent/services/agent_helpers/tashkeel_agent.py

+import logging
+from services.agent_helpers.agent_prompts import tashkeel_agent_prompt
+
+logger = logging.getLogger(__name__)
+
+class TashkeelAgent:
+    """Agent to apply Arabic (Egyptian) diacritization on text for TTS"""
+
+    def __init__(self, openai_service):
+        self.openai_service = openai_service
+
+    def apply_tashkeel(self, text: str) -> str:
+        """Send text to LLM and return fully diacritized version"""
+        try:
+            if not self.openai_service.is_available():
+                logger.warning("OpenAI service not available for TashkeelAgent")
+                return text  # fallback: return original
+
+            messages = [
+                {"role": "system", "content": tashkeel_agent_prompt},
+                {"role": "user", "content": text}
+            ]
+            response = self.openai_service.client.chat.completions.create(
+                model="gpt-4o-mini",  # أو أي موديل خفيف سريع
+                messages=messages,
+                temperature=0.1
+            )
+            return response.choices[0].message.content.strip()
+        except Exception as e:
+            logger.error(f"TashkeelAgent error: {e}")
+            return text  # fallback

self_hosted_env/voice_agent/services/agent_helpers/tashkeel_fixer.py

+custom_fixes = {
+    "التكيف": "التَكَيُّف",
+    "بقاء": "البَقَّاء",
+    "قدرة": "القُدرَة",
+    "النقل": "النَقْل",
+    "الدب": "الدُبّ",
+    "النمر": "النَمِر",
+    "فرو": "فَروُ",
+    "البني": "البُنّي",
+    "ملونة": "مِلوِنَةْ",
+    "قوس قزح": "قُوس قُزَح",
+    "معينة": "مُعيَّنَة",
+    "الفنك": "الفنِك",
+    "الحر": "الحَر",
+    "الشم": "الَشَمْ",
+    "البصر": "البَصَر",
+    "الأذن": "الاُذُن",
+    "الفم": "الفَم",
+    "العين": "العِين",
+    "اللهث": "اللَّهْث",
+    "القطط": "القطط",
+    "لنقل": "لنَقْل",
+    "قدم": "قَدَمْ",
+    "مية": "مَيَّةْ",
+    "حاسة": "حاسة",
+}
+
+
+
+def apply_fixes(text, fixes_dict):
+    for wrong, fixed in fixes_dict.items():
+        if wrong in text:
+            text = text.replace(wrong, fixed)
+    return text

self_hosted_env/voice_agent/services/agent_service.py

@@ -14,6 +14,8 @@ from services.connection_pool import ConnectionPool
 from services.agent_helpers.query_handlers import QueryHandler
 from services.agent_helpers.context_generator import ContextGenerator
 from services.agent_helpers.response_generator import ResponseGenerator
+from services.agent_helpers.tashkeel_agent import TashkeelAgent
+from services.agent_helpers.tashkeel_fixer import apply_fixes, custom_fixes
 from services.tts.tts_manager import get_tts_service
 
 logger = logging.getLogger(__name__)
@@ -64,20 +66,34 @@ class AgentService:
             self.query_handler, self.context_generator
         )
 
+        self.tashkeel_agent = TashkeelAgent(self.openai_service)
+
+
     def is_available(self) -> bool:
         return self.openai_service.is_available()
 
     def text_to_speech(self, text: str, language: str) -> bytes:
         if not self.tts_service or not self.tts_service.is_available():
             raise HTTPException(status_code=503, detail="TTS service is not available")
+
+        # Step 1: apply tashkeel before sending to TTS
+        text = self.tashkeel_agent.apply_tashkeel(text)
+        print(f"Tashkeel applied: {text}")
+
+        # Step 2: send to TTS
         return self.tts_service.generate_speech(text, language)
 
+
     def generate_response(self, user_message: str, student_id: str, subject: str = "Science", 
                          model: str = Models.chat, temperature: float = 0.3, top_k: int = 3) -> str:
         """Main response generation method"""
-        return self.response_generator.generate_response(
+        response = self.response_generator.generate_response(
             user_message, student_id, subject, model, temperature, top_k
         )
+        response = apply_fixes(response, custom_fixes)
+        #response = self.tashkeel_agent.apply_tashkeel(response)
+        print(f"response: {response}")
+        return response
 
     def search_similar(self, query_embedding: List[float], student_id: str, 
                       subject: str = "chemistry", top_k: int = 3):

self_hosted_env/voice_agent/services/segmentation_service.py

@@ -6,10 +6,10 @@ class LanguageSegmentationService:
     A service to segment a string of text into a list of dictionaries,
     each tagged with its detected language.
     """
+    
     def segment_text(self, text: str) -> List[Dict[str, str]]:
         """
         Takes a mixed-language string and splits it into segments.
-
         Example:
             Input: "هذا هو a test of the system."
             Output: [
@@ -32,7 +32,13 @@ class LanguageSegmentationService:
         for word in words:
             word_lang = self._detect_word_language(word)
             
-            if word_lang == current_lang:
+            # Check if this is a "neutral" token (numbers, punctuation, special markers)
+            is_neutral = self._is_neutral_token(word)
+            
+            if is_neutral:
+                # Neutral tokens stay with the current segment
+                current_segment.append(word)
+            elif word_lang == current_lang:
                 # If the language is the same, add the word to the current segment
                 current_segment.append(word)
             else:
@@ -42,7 +48,6 @@ class LanguageSegmentationService:
                         "text": " ".join(current_segment),
                         "language": current_lang
                     })
-                
                 # Start a new segment with the new word and language
                 current_lang = word_lang
                 current_segment = [word]
@@ -57,11 +62,32 @@ class LanguageSegmentationService:
         print(f"Segmented text into {len(segments)} parts.")
         return segments
     
+    def _is_neutral_token(self, word: str) -> bool:
+        """
+        Check if a token is 'neutral' (numbers, punctuation, special markers).
+        These should stick with the current segment rather than create a new one.
+        """
+        # Strip common punctuation to check the core content
+        stripped = word.strip('.,!?;:()[]{}"\'-')
+        
+        # Empty after stripping (pure punctuation)
+        if not stripped:
+            return True
+        
+        # Pure numbers (with optional punctuation like "1." or "#1")
+        if stripped.replace('#', '').isdigit():
+            return True
+        
+        # Special markdown-like markers (##, ###, etc.)
+        if all(c == '#' for c in stripped):
+            return True
+        
+        return False
+    
     def _detect_word_language(self, word: str) -> str:
         """Detects language of a single word, defaulting to 'en' for ambiguity."""
         # Simple heuristic: if it contains any Arabic characters, it's Arabic.
         if any('\u0600' <= char <= '\u06FF' for char in word):
             return "ar"
-        
         # For non-Arabic words, we can assume English
         return "en"
\ No newline at end of file