Update app.py

app.py

@@ -1,18 +1,21 @@
 import gradio as gr
 import torch
-from transformers import pipeline, AutoTokenizer, AutoModelForSeq2SeqLM
+from transformers import pipeline
+import requests
 import re
+import os
+from huggingface_hub import login
 
-# Global variables to store the models
+# Authenticate with Hugging Face
+if "HF_TOKEN" in os.environ:
+    login(token=os.environ["HF_TOKEN"])
+
+# Global variable to store the Atlas-Chat model
 atlas_pipe = None
-transliteration_tokenizer = None
-transliteration_model = None
 
-def load_models():
-    """Load both Atlas-Chat and Transliteration models"""
-    global atlas_pipe, transliteration_tokenizer, transliteration_model
-    
-    # Load Atlas-Chat model
+def load_atlas_model():
+    """Load only the Atlas-Chat model locally"""
+    global atlas_pipe
     if atlas_pipe is None:
         print("🏔️ Loading Atlas-Chat-2B model...")
         atlas_pipe = pipeline(
@@ -22,15 +25,7 @@ def load_models():
             device="cuda" if torch.cuda.is_available() else "cpu"
         )
         print("✅ Atlas-Chat model loaded!")
-    
-    # Load Transliteration model
-    if transliteration_tokenizer is None or transliteration_model is None:
-        print("🔄 Loading Transliteration model...")
-        transliteration_tokenizer = AutoTokenizer.from_pretrained("atlasia/Transliteration-Moroccan-Darija")
-        transliteration_model = AutoModelForSeq2SeqLM.from_pretrained("atlasia/Transliteration-Moroccan-Darija")
-        print("✅ Transliteration model loaded!")
-    
-    return atlas_pipe, transliteration_tokenizer, transliteration_model
+    return atlas_pipe
 
 def detect_arabizi(text):
     """
@@ -89,40 +84,70 @@ def detect_arabizi(text):
     
     return False
 
-def arabizi_to_arabic_ai(arabizi_text):
+def arabizi_to_arabic_api(arabizi_text):
     """
-    Convert Arabizi text to Arabic using the specialized AI model
+    Convert Arabizi text to Arabic using Hugging Face Inference API
     """
     try:
-        _, tokenizer, model = load_models()
+        # Check if HF_TOKEN is available
+        if "HF_TOKEN" not in os.environ:
+            print("❌ HF_TOKEN not found, falling back to original text")
+            return arabizi_text
         
-        # Tokenize the input text
-        input_tokens = tokenizer(arabizi_text, return_tensors="pt", padding=True, truncation=True, max_length=512)
+        API_URL = "https://api-inference.huggingface.co/models/atlasia/Transliteration-Moroccan-Darija"
+        headers = {"Authorization": f"Bearer {os.environ['HF_TOKEN']}"}
         
-        # Perform transliteration
-        with torch.no_grad():
-            output_tokens = model.generate(
-                **input_tokens,
-                max_length=512,
-                num_beams=4,
-                early_stopping=True,
-                no_repeat_ngram_size=2
-            )
+        # Prepare the payload
+        payload = {
+            "inputs": arabizi_text,
+            "parameters": {
+                "max_length": 512,
+                "num_beams": 4,
+                "early_stopping": True
+            }
+        }
         
-        # Decode the output tokens
-        arabic_text = tokenizer.decode(output_tokens[0], skip_special_tokens=True)
-        
-        return arabic_text.strip()
+        # Make API request with timeout
+        response = requests.post(API_URL, headers=headers, json=payload, timeout=30)
         
+        # Check if request was successful
+        if response.status_code == 200:
+            result = response.json()
+            
+            # Handle different response formats
+            if isinstance(result, list) and len(result) > 0:
+                if "generated_text" in result[0]:
+                    return result[0]["generated_text"].strip()
+                elif isinstance(result[0], str):
+                    return result[0].strip()
+            elif isinstance(result, dict) and "generated_text" in result:
+                return result["generated_text"].strip()
+            elif isinstance(result, str):
+                return result.strip()
+            else:
+                print(f"❌ Unexpected API response format: {result}")
+                return arabizi_text
+                
+        elif response.status_code == 503:
+            print("⏳ Model is loading, falling back to original text")
+            return arabizi_text
+        else:
+            print(f"❌ API error {response.status_code}: {response.text}")
+            return arabizi_text
+            
+    except requests.exceptions.Timeout:
+        print("⏰ API timeout, falling back to original text")
+        return arabizi_text
+    except requests.exceptions.RequestException as e:
+        print(f"❌ API request failed: {e}")
+        return arabizi_text
     except Exception as e:
-        print(f"❌ Error in Arabizi→Arabic conversion: {e}")
-        # Fallback to original text if conversion fails
+        print(f"❌ Unexpected error in API conversion: {e}")
         return arabizi_text
 
 def arabic_to_arabizi(arabic_text):
     """
-    Convert Arabic script to Arabizi using character mappings
-    (Keeping this as backup since no reverse model available)
+    Convert Arabic script to Arabizi using comprehensive hard-coded mappings
     """
     if not arabic_text:
         return arabic_text
@@ -151,35 +176,17 @@ def arabic_to_arabizi(arabic_text):
         'طاجين': 'tajine', 'أتاي': 'atay', 'خوبز': 'khobz',
         'كاين': 'kayn', 'ماكاينش': 'makaynsh', 'شي': 'chi',
         'زوين': 'zwin', 'زوينا': 'zwina', 'مزيان': 'mzyan', 'مزيانا': 'mzyana',
-        'درت': 'dert', 'درتي': 'derti', 'دار': 'dar', 'درات': 'derat',
-        'مشيت': 'mchit', 'مشيتي': 'mchiti', 'مشا': 'mcha', 'مشات': 'mchat',
-        'جيت': 'jit', 'جيتي': 'jiti', 'جا': 'ja', 'جات': 'jat',
-        'شفت': 'cheft', 'شفتي': 'chefti', 'شاف': 'chaf', 'شافت': 'chafat',
-        'سمعت': 'sme3t', 'سمعتي': 'sme3ti', 'سمع': 'sma3', 'سمعات': 'sma3at',
-        'أكلت': 'klit', 'أكلتي': 'kliti', 'كلا': 'kla', 'كلات': 'klat',
-        'شربت': 'chrebt', 'شربتي': 'chrebti', 'شرب': 'chreb', 'شربات': 'chrebat',
-        'نعست': 'ne3st', 'نعستي': 'ne3sti', 'نعس': 'ne3s', 'نعسات': 'ne3sat',
-        'خرجت': 'khrjt', 'خرجتي': 'khrjti', 'خرج': 'khrj', 'خرجات': 'khrjat',
-        'دخلت': 'dkhlt', 'دخلتي': 'dkhlti', 'دخل': 'dkhl', 'دخلات': 'dkhlat',
-        'قريت': 'qrit', 'قريتي': 'qriti', 'قرا': 'qra', 'قرات': 'qrat',
-        'كتبت': 'ktebt', 'كتبتي': 'ktebti', 'كتب': 'kteb', 'كتبات': 'ktebat',
-        'لعبت': 'l3ebt', 'لعبتي': 'l3ebti', 'لعب': 'l3eb', 'لعبات': 'l3ebat',
-        'خدمت': 'khdmt', 'خدمتي': 'khdmti', 'خدم': 'khdm', 'خدمات': 'khdmat',
-        'صليت': 'sllit', 'صليتي': 'slliti', 'صلا': 'slla', 'صلات': 'sllat',
-        'طبخت': '6bkht', 'طبختي': '6bkhti', 'طبخ': '6bekh', 'طبخات': '6bekhat',
-        'واحد': 'wa7ed', 'جوج': 'joj', 'تلاتا': 'tlata', 'ربعا': 'reb3a',
-        'خمسا': 'khamsa', 'ستا': 'setta', 'سبعا': 'seb3a', 'تمنيا': 'tmnya',
-        'تسعا': 'tes3a', 'عشرا': '3echra', 'حداش': '7dach', 'طناش': '6nach',
-        'نهار': 'nhar', 'ليلا': 'lila', 'صباح': 'sba7', 'عشيا': '3echiya',
-        'أمس': 'ems', 'البارح': 'lbare7', 'غدا': 'ghda', 'بعد غدا': 'b3d ghda',
-        'دار': 'dar', 'بيت': 'bit', 'شارع': 'char3', 'مدينا': 'mdina',
-        'كرهوبا': 'karhouba', 'طوموبيل': 'tomobil', 'قطار': 'q6ar', 'باص': 'bas',
-        'ماكلا': 'makla', 'شراب': 'chrab', 'لما': 'lma', 'عطش': '36ch',
-        'جوع': 'jo3', 'شبعان': 'cheb3an', 'عيان': '3yyan', 'صحيح': 's7i7',
-        'مريض': 'mrid', 'دكتور': 'doktor', 'سبيطار': 'sbitar', 'دوا': 'dwa',
-        'فلوس': 'flous', 'درهم': 'derhem', 'ريال': 'riyal', 'اليورو': 'lyoro',
-        'خدما': 'khedma', 'معلم': 'mo3alim', 'طالب': 'talib', 'أستاذ': 'ostaz',
-        'كتاب': 'ktab', 'قلم': 'qalam', 'كاغط': 'kaghet', 'طاولا': 'tabla'
+        'كاينين': 'kaynin', 'مطعم': 'ma63am', 'مطاعم': 'ma6a3im',
+        'مشهور': 'mashhur', 'مشهورين': 'mashhurin', 'وسط': 'wost',
+        'المدينة': 'lmdina', 'مدينة': 'mdina', 'إيطالي': 'italiy',
+        'ياباني': 'yabani', 'مغربي': 'maghribi', 'فرنسي': 'fransi',
+        'أمريكي': 'amriki', 'صيني': 'sini', 'هندي': 'hindi',
+        'لحم': 'la7m', 'دجاج': 'djaj', 'حوت': '7ut', 'خضرة': 'khodra',
+        'فواكه': 'fawakeh', 'جبن': 'jben', 'زبدة': 'zebda', 'حليب': '7lib',
+        'قهوة': 'qahwa', 'شاي': 'atay', 'ماء': 'ma', 'عصير': '3asir',
+        'خبز': 'khobz', 'رز': 'roz', 'مكرونة': 'makarona', 'بطاطا': 'batata',
+        'طماطم': 'toma6im', 'بصل': 'basal', 'ثوم': 'tum', 'فلفل': 'felfel',
+        'ملح': 'mel7', 'سكر': 'sokkar', 'زيت': 'zit', 'خل': 'khall'
     }
     
     # CHARACTER MAPPINGS (Arabic → Arabizi)
@@ -196,7 +203,7 @@ def arabic_to_arabizi(arabic_text):
     
     result = arabic_text
     
-    # Step 1: Apply word mappings
+    # Step 1: Apply word mappings (most specific first)
     for arabic_word, arabizi_word in word_mappings.items():
         # Use word boundaries to avoid partial matches
         result = re.sub(r'\b' + re.escape(arabic_word) + r'\b', arabizi_word, result)
@@ -208,39 +215,36 @@ def arabic_to_arabizi(arabic_text):
     return result.strip()
 
 def chat_with_atlas(message, history):
-    """Generate response from Atlas-Chat model with AI-powered Arabizi conversion"""
+    """Generate response from Atlas-Chat model with API-powered Arabizi conversion"""
     if not message.strip():
         return "ahlan wa sahlan! kifash n9der n3awnek? / مرحبا! كيفاش نقدر نعاونك؟"
     
     try:
-        # Load models
-        atlas_model, _, _ = load_models()
+        # Load Atlas-Chat model
+        atlas_model = load_atlas_model()
         
         # Detect if input is Arabizi
         is_arabizi_input = detect_arabizi(message)
         
-        print("\n" + "="*60)
-        print("🔍 DEBUG LOG - FULL CONVERSION PIPELINE")
-        print("="*60)
-        print(f"📥 ORIGINAL INPUT: '{message}'")
-        print(f"🤖 ARABIZI DETECTED: {is_arabizi_input}")
+        print("\n" + "="*50)
+        print("🔍 ATLAS-CHAT DEBUG LOG")
+        print("="*50)
+        print(f"📥 INPUT: '{message}'")
+        print(f"🔍 ARABIZI: {is_arabizi_input}")
         
         # Prepare input for the model
         if is_arabizi_input:
-            # Convert Arabizi to Arabic using AI model
-            print(f"\n🔄 STEP 1: Converting Arabizi to Arabic...")
-            arabic_input = arabizi_to_arabic_ai(message)
-            print(f"✅ ARABIC CONVERSION: '{arabic_input}'")
+            print("🔄 Converting Arabizi→Arabic via API...")
+            arabic_input = arabizi_to_arabic_api(message)
+            print(f"✅ ARABIC: '{arabic_input}'")
             model_input = arabic_input
         else:
-            # Use original input (Arabic or English)
-            print(f"\n➡️  NO CONVERSION NEEDED - Using original input")
+            print("➡️  No conversion needed")
             model_input = message
         
-        print(f"\n🤖 STEP 2: Sending to Atlas-Chat model...")
-        print(f"📤 MODEL INPUT: '{model_input}'")
+        print(f"🤖 Sending to Atlas-Chat...")
         
-        # Generate response using Arabic input
+        # Generate response using Atlas-Chat
         messages = [{"role": "user", "content": model_input}]
         
         outputs = atlas_model(
@@ -253,28 +257,22 @@ def chat_with_atlas(message, history):
         
         # Extract the response
         response = outputs[0]["generated_text"][-1]["content"].strip()
-        print(f"✅ MODEL RESPONSE (Arabic): '{response}'")
+        print(f"✅ RESPONSE: '{response[:100]}{'...' if len(response) > 100 else ''}'")
         
         # Convert response back to Arabizi if input was Arabizi
         if is_arabizi_input:
-            print(f"\n🔄 STEP 3: Converting response back to Arabizi...")
+            print("🔄 Converting Arabic→Arabizi...")
             arabizi_response = arabic_to_arabizi(response)
-            print(f"✅ FINAL ARABIZI RESPONSE: '{arabizi_response}'")
-            print("="*60)
-            print("🎯 FINAL OUTPUT TO USER:", arabizi_response)
-            print("="*60 + "\n")
+            print(f"✅ FINAL: '{arabizi_response[:100]}{'...' if len(arabizi_response) > 100 else ''}'")
+            print("="*50 + "\n")
             return arabizi_response
         else:
-            # Return original response for Arabic/English
-            print(f"\n➡️  NO BACK-CONVERSION NEEDED")
-            print("="*60)
-            print("🎯 FINAL OUTPUT TO USER:", response)
-            print("="*60 + "\n")
+            print("="*50 + "\n")
             return response
         
     except Exception as e:
-        print(f"\n❌ ERROR OCCURRED: {str(e)}")
-        print("="*60 + "\n")
+        print(f"\n❌ ERROR: {str(e)}")
+        print("="*50 + "\n")
         # Return error in appropriate language
         if detect_arabizi(message):
             return f"sorry, kan chi mochkil: {str(e)}. 3awd jar'b!"
@@ -284,19 +282,20 @@ def chat_with_atlas(message, history):
 # Create the Gradio interface
 demo = gr.ChatInterface(
     fn=chat_with_atlas,
-    title="🏔️ Atlas-Chat: Advanced Moroccan Arabic AI",
+    title="🏔️ Atlas-Chat: AI-Powered Moroccan Arabic Assistant",
     description="""
-    **مرحبا بك في أطلس شات المطور!** Welcome to Advanced Atlas-Chat! 🇲🇦
+    **مرحبا بك في أطلس شات!** Welcome to Atlas-Chat! 🇲🇦
     
-    **🧠 AI-Powered Language Detection & Conversion:**
-    - **Arabic Script (العربية)** → AI responds in Arabic
-    - **Arabizi (3arabi bi 7oruf latin)** → AI-powered conversion → Arabizi response  
-    - **English** → AI responds in English
+    **🚀 Powered by Hugging Face Inference API:**
+    - **Arabic Script (العربية)** → Direct conversation
+    - **Arabizi (3arabi bi 7oruf latin)** → API conversion → Arabizi response  
+    - **English** → Direct conversation
     
-    **⚡ Professional Arabizi Conversion**
-    - Uses specialized AI model trained on Moroccan Darija
-    - Perfect understanding of context: "kayn chi" → "كاين شي"
-    - Handles complex phrases accurately
+    **⚡ Features:**
+    - Professional AI Arabizi conversion via API
+    - No local model conflicts
+    - Fast and reliable responses
+    - Comprehensive language detection
     
     **جرب هذه الأسئلة / Try these questions:**
     """,