chatbot-salim

app.py

@@ -1,734 +1,174 @@
 import gradio as gr
 import os
+from sentence_transformers import SentenceTransformer
 from qdrant_client import QdrantClient
 import requests
 from groq import Groq
 import json
-import re
-import unicodedata
 from datetime import datetime
-from typing import Tuple, List, Dict, Optional, Any
 
 
 # --- Configuration ---
+DEEPSEEK_API_KEY = os.getenv('DEEPSEEK_API_KEY')
 QDRANT_URL = "https://f1db4954-34db-4b00-b9bb-8acf74c1d6e4.eu-central-1-0.aws.cloud.qdrant.io"
 QDRANT_API_KEY = os.getenv('QDRANT_API_KEY')
 GROQ_API_KEY = os.getenv("GROQ_API_KEY")
-VOYAGE_API_KEY = os.getenv("VOYAGE_API_KEY")
 
-DB_Server = os.getenv("DB_Server")
-DB_User = os.getenv("DB_User")
-DB_Password = os.getenv("DB_Password")
-DB_Name = os.getenv("DB_Name")
+print("🔑 DEEPSEEK_API_KEY:", DEEPSEEK_API_KEY)
+print("🔑 QDRANT_API_KEY:", QDRANT_API_KEY)
+print("🔑 GROQ_API_KEY:", GROQ_API_KEY)
+print("🔧 Gradio version:", gr.__version__)
 
+# --- Initialize Clients ---
+print("🚀 Initializing RAG Chatbot...")
 
-# =============================================================================
-# ADVANCED TOPIC CLASSIFICATION & FILTERING
-# =============================================================================
-
-class TopicClassifier:
-    """
-    Advanced topic classifier to ensure the chatbot stays on-topic.
-    Uses keyword matching + semantic understanding to filter off-topic queries.
-    """
-    
-    # Allowed domains - Real estate in Algeria
-    ALLOWED_TOPICS = {
-        "real_estate_search": [
-            "appartement", "maison", "terrain", "villa", "studio", "duplex", "local",
-            "f1", "f2", "f3", "f4", "f5", "f6", "logement", "habitation", "residence",
-            "immeuble", "etage", "chambre", "salon", "cuisine", "salle de bain",
-            "garage", "parking", "balcon", "terrasse", "jardin", "cave", "grenier",
-            "surface", "superficie", "m2", "m²", "pièces", "pieces"
-        ],
-        "real_estate_transaction": [
-            "acheter", "achat", "vendre", "vente", "louer", "location", "bail",
-            "locataire", "propriétaire", "proprietaire", "bailleur", "prix", "budget",
-            "négocier", "negocier", "offre", "contre-offre", "compromis", "promesse",
-            "signature", "acte", "notaire", "frais", "commission", "honoraires",
-            "acompte", "caution", "garantie", "depot", "dépôt", "loyer", "charges"
-        ],
-        "legal_admin": [
-            "document", "papier", "dossier", "formalité", "formalite", "procédure",
-            "procedure", "démarche", "demarche", "administration", "mairie", "préfecture",
-            "cadastre", "conservation", "hypothèque", "hypotheque", "enregistrement",
-            "timbre", "taxe", "impôt", "impot", "fiscalité", "fiscalite", "tva",
-            "livret foncier", "acte de propriété", "titre foncier", "permis de construire",
-            "certificat", "urbanisme", "zonage", "loi", "réglementation", "reglementation",
-            "droit", "juridique", "légal", "legal", "contrat", "clause"
-        ],
-        "financing": [
-            "crédit", "credit", "prêt", "pret", "emprunt", "banque", "financement",
-            "taux", "intérêt", "interet", "mensualité", "mensualite", "apport",
-            "capacité", "capacite", "endettement", "simulation", "simulateur",
-            "assurance", "garantie", "hypothécaire", "hypothecaire", "remboursement"
-        ],
-        "market_info": [
-            "marché", "marche", "immobilier", "investissement", "investir", "rentabilité",
-            "rentabilite", "rendement", "prix moyen", "estimation", "évaluation",
-            "evaluation", "expertise", "tendance", "évolution", "evolution", "quartier",
-            "zone", "secteur", "emplacement", "localisation"
-        ],
-        "algerian_cities": [
-            "alger", "oran", "constantine", "annaba", "blida", "bejaia", "béjaïa",
-            "tizi ouzou", "setif", "sétif", "batna", "chlef", "boumerdes", "boumerdès",
-            "mostaganem", "relizane", "tlemcen", "sidi bel abbes", "béchar", "bechar",
-            "skikda", "ghardaia", "guelma", "tiaret", "msila", "m'sila", "el oued",
-            "ouargla", "tipaza", "djelfa", "naama", "médea", "medea", "saida", "saïda",
-            "khenchela", "el bayadh", "illizi", "ain temouchent", "aïn temouchent",
-            "bordj bou arreridj", "bouira", "algerie", "algérie", "dz"
-        ],
-        "greetings": [
-            "bonjour", "bonsoir", "salut", "hello", "hi", "hey", "coucou", "salam",
-            "bonne journée", "bonne soirée", "comment allez", "comment vas", "ça va",
-            "ca va", "merci", "au revoir", "à bientôt", "a bientot"
-        ]
-    }
-    
-    # Explicitly OFF-TOPIC keywords
-    OFF_TOPIC_KEYWORDS = [
-        # General off-topic
-        "recette", "cuisine", "manger", "restaurant", "football", "sport", "match",
-        "film", "musique", "chanson", "jeu", "game", "video", "politique", "élection",
-        "president", "gouvernement", "religion", "météo", "meteo", "vacances", "voyage",
-        "avion", "train", "voiture", "moto", "vélo", "santé", "sante", "médecin",
-        "hôpital", "hopital", "maladie", "covid", "école", "ecole", "université",
-        "universite", "examen", "cours", "professeur", "étudiant", "etudiant",
-        
-        # Tech unrelated
-        "programmer", "coder", "javascript", "python", "java", "android", "iphone",
-        "ordinateur", "laptop", "téléphone", "telephone", "internet", "wifi",
-        
-        # Completely unrelated
-        "guerre", "armée", "armee", "militaire", "drogue", "arme", "violence",
-        "sexe", "pornographie", "gambling", "casino", "pari", "crypto", "bitcoin",
-        
-        # Other services
-        "coiffeur", "dentiste", "avocat", "médecin", "docteur", "plombier",
-        "électricien", "electricien", "mécanicien", "mecanicien"
-    ]
-    
-    @classmethod
-    def normalize(cls, text: str) -> str:
-        if not text:
-            return ""
-        text = text.lower()
-        text = unicodedata.normalize('NFD', text)
-        text = re.sub(r"[\u0300-\u036f]", "", text)
-        text = re.sub(r"\s+", " ", text)
-        return text.strip()
-    
-    @classmethod
-    def classify(cls, query: str) -> Tuple[bool, str, float]:
-        """
-        Classify if query is on-topic for real estate.
-        Returns: (is_on_topic, detected_category, confidence_score)
-        """
-        normalized = cls.normalize(query)
-        words = set(normalized.split())
-        
-        # Check for explicit off-topic keywords
-        off_topic_matches = sum(1 for kw in cls.OFF_TOPIC_KEYWORDS if kw in normalized)
-        if off_topic_matches >= 2:
-            return False, "off_topic", 0.9
-        
-        # Check for on-topic keywords
-        topic_scores = {}
-        for category, keywords in cls.ALLOWED_TOPICS.items():
-            score = sum(1 for kw in keywords if kw in normalized)
-            if score > 0:
-                topic_scores[category] = score
-        
-        if topic_scores:
-            best_category = max(topic_scores, key=topic_scores.get)
-            confidence = min(topic_scores[best_category] / 3.0, 1.0)
-            return True, best_category, confidence
-        
-        # Ambiguous - check query length and patterns
-        if len(words) <= 3:
-            # Short queries might be greetings or simple questions
-            return True, "general", 0.3
-        
-        # Default: slightly lean towards allowing (benefit of doubt)
-        return True, "unknown", 0.2
-    
-    @classmethod
-    def get_off_topic_response(cls) -> str:
-        return (
-            "🏠 Je suis un assistant spécialisé exclusivement dans **l'immobilier en Algérie**.\n\n"
-            "Je peux vous aider avec :\n"
-            "• Recherche de biens (appartements, maisons, terrains, villas...)\n"
-            "• Informations sur l'achat, la vente ou la location\n"
-            "• Démarches administratives et juridiques\n"
-            "• Financement et crédits immobiliers\n"
-            "• Prix et tendances du marché algérien\n\n"
-            "Votre question semble hors de mon domaine d'expertise. "
-            "Pourriez-vous reformuler votre demande en rapport avec l'immobilier ?"
-        )
-
-
-# =============================================================================
-# INTENT DETECTION - More granular
-# =============================================================================
-
-class IntentDetector:
-    """
-    Detect user intent for more precise responses.
-    """
-    
-    INTENTS = {
-        "greeting": {
-            "patterns": ["bonjour", "bonsoir", "salut", "hello", "hi", "hey", "coucou", "salam"],
-            "max_words": 5
-        },
-        "thanks": {
-            "patterns": ["merci", "thank", "remercie"],
-            "max_words": 10
-        },
-        "goodbye": {
-            "patterns": ["au revoir", "bye", "à bientôt", "a bientot", "bonne journée", "bonne soirée"],
-            "max_words": 6
-        },
-        "property_search": {
-            "patterns": ["cherche", "recherche", "trouve", "trouver", "besoin", "veux", "voudrais", 
-                        "souhaite", "intéressé", "interesse", "disponible"],
-            "requires": ["appartement", "maison", "villa", "terrain", "studio", "f1", "f2", "f3", 
-                        "f4", "f5", "duplex", "local", "logement"]
-        },
-        "price_inquiry": {
-            "patterns": ["prix", "coût", "cout", "combien", "budget", "tarif", "estimation"],
-            "requires": []
-        },
-        "procedure_inquiry": {
-            "patterns": ["comment", "étape", "etape", "procédure", "procedure", "démarche", 
-                        "demarche", "document", "formalité", "formalite", "quoi faire"],
-            "requires": []
-        },
-        "legal_inquiry": {
-            "patterns": ["loi", "légal", "legal", "droit", "juridique", "réglementation", 
-                        "reglementation", "permis", "autorisation"],
-            "requires": []
-        },
-        "comparison": {
-            "patterns": ["différence", "difference", "comparer", "comparaison", "mieux", 
-                        "meilleur", "vs", "ou bien", "lequel"],
-            "requires": []
-        }
-    }
-    
-    @classmethod
-    def detect(cls, query: str) -> Tuple[str, float]:
-        """
-        Detect the primary intent of the query.
-        Returns: (intent_name, confidence)
-        """
-        normalized = TopicClassifier.normalize(query)
-        word_count = len(normalized.split())
-        
-        for intent, config in cls.INTENTS.items():
-            patterns = config.get("patterns", [])
-            max_words = config.get("max_words", 100)
-            requires = config.get("requires", [])
-            
-            pattern_match = any(p in normalized for p in patterns)
-            
-            if pattern_match:
-                if max_words and word_count > max_words:
-                    continue
-                if requires and not any(r in normalized for r in requires):
-                    continue
-                return intent, 0.8
-        
-        return "general_query", 0.5
-
-
-# =============================================================================
-# RESPONSE TEMPLATES
-# =============================================================================
-
-class ResponseTemplates:
-    """
-    Pre-built responses for common intents.
-    """
-    
-    GREETING = (
-        "👋 Bonjour et bienvenue !\n\n"
-        "Je suis votre assistant immobilier spécialisé en **Algérie**. "
-        "Je suis là pour vous accompagner dans vos projets immobiliers.\n\n"
-        "**Comment puis-je vous aider ?**\n"
-        "• Rechercher un bien (appartement, maison, terrain...)\n"
-        "• Répondre à vos questions sur l'achat ou la location\n"
-        "• Expliquer les démarches administratives\n"
-        "• Vous informer sur les prix du marché\n\n"
-        "N'hésitez pas à me poser votre question ! 😊"
-    )
-    
-    THANKS = (
-        "De rien ! Je suis ravi d'avoir pu vous aider. 😊\n\n"
-        "N'hésitez pas si vous avez d'autres questions sur l'immobilier en Algérie."
-    )
-    
-    GOODBYE = (
-        "Au revoir et bonne continuation dans vos recherches ! 👋\n\n"
-        "N'hésitez pas à revenir si vous avez d'autres questions immobilières."
-    )
-    
-    NO_RESULTS = (
-        "Je n'ai pas trouvé d'informations correspondant exactement à votre demande "
-        "dans notre base de données.\n\n"
-        "**Suggestions :**\n"
-        "• Essayez avec des critères différents (ville, type de bien, budget)\n"
-        "• Consultez notre liste complète de biens : https://jacheteenalgerie.com/liste-des-biens/\n"
-        "• Prenez rendez-vous avec un expert : https://jacheteenalgerie.com/prendre-rdv/"
-    )
-
-
-# =============================================================================
-# IMPROVED EMBEDDING & SEARCH
-# =============================================================================
-
-print("🚀 Initializing Advanced RAG Chatbot...")
-
-def get_embedding_via_voyage_ai(text: str) -> Optional[List[float]]:
-    """Get embeddings with error handling and retry logic."""
-    url = "https://api.voyageai.com/v1/embeddings"
-    headers = {
-        "Authorization": f"Bearer {VOYAGE_API_KEY}",
-        "Content-Type": "application/json"
-    }
-    payload = {
-        "input": [text],
-        "model": "voyage-2"
-    }
-    
-    for attempt in range(3):
-        try:
-            response = requests.post(url, headers=headers, json=payload, timeout=30)
-            response.raise_for_status()
-            result = response.json()
-            if result and "data" in result and len(result["data"]) > 0:
-                return result["data"][0]["embedding"]
-        except requests.exceptions.RequestException as e:
-            print(f"⚠️ Voyage AI attempt {attempt + 1} failed: {e}")
-            if attempt == 2:
-                return None
-    return None
-
+# Initialize embedding model
+print("📡 Loading embedding model...")
+embedding_model = SentenceTransformer('paraphrase-MiniLM-L3-v2')
+print("✅ Embedding model loaded!")
 
 # Initialize Qdrant client
 print("🔗 Connecting to Qdrant...")
-qdrant_client = None
 try:
-    qdrant_client = QdrantClient(
-        url=QDRANT_URL,
-        api_key=QDRANT_API_KEY,
-        check_compatibility=False,
-        prefer_grpc=False  # Use REST API to avoid asyncio event loop issues
-    )
-    collection_info = qdrant_client.get_collection(collection_name="biens")
-    print(f"✅ Connected to Qdrant! Collection 'biens' has {collection_info.points_count} documents.")
-    collection_info_articles = qdrant_client.get_collection(collection_name="articles")
-    print(f"✅ Connected to Qdrant! Collection 'articles' has {collection_info_articles.points_count} documents.")
+    qdrant_client = QdrantClient(url=QDRANT_URL, api_key=QDRANT_API_KEY)
+    collection_info = qdrant_client.get_collection(collection_name="RAG JEA")
+    print(f"✅ Connected to Qdrant! Collection 'RAG JEA' has {collection_info.points_count} documents.")
 except Exception as e:
     print(f"❌ Failed to connect to Qdrant: {e}")
+    qdrant_client = None
 
 # Initialize Groq client
 groq_client = Groq(api_key=GROQ_API_KEY) if GROQ_API_KEY else None
 
 
-# =============================================================================
-# LOGGING & DATABASE
-# =============================================================================
-
-def log_interaction(user_message: str, response: str, context: str, 
-                   tokens_max: int, temperature: float, top_p: float):
+def log_interaction(user_message, response, context, tokens_max, temperature, top_p):
     log_entry = {
         "timestamp": datetime.utcnow().isoformat(),
         "question": user_message,
         "response": response,
-        "context": context[:500] if context else "",
+        "context": context,
         "tokens_max": tokens_max,
         "temperature": temperature,
         "top_p": top_p,
     }
     try:
-        requests.post("https://ee98552a720c.ngrok-free.app/log", json=log_entry, timeout=5)
+        requests.post("https://ee98552a720c.ngrok-free.app/log", json=log_entry)
     except Exception as e:
-        print(f"⚠️ Failed to send log: {e}")
-
-
-def save_chat_to_mysql(question: str, answer: str, source: str):
-    """Save Q&A to MySQL database."""
-    try:
-        if not all([DB_Server, DB_User, DB_Password, DB_Name]):
-            return
-        
-        import pymysql
-        connection = pymysql.connect(
-            host=DB_Server,
-            user=DB_User,
-            password=DB_Password,
-            database=DB_Name,
-            charset="utf8mb4",
-            cursorclass=pymysql.cursors.DictCursor,
-            autocommit=True,
-        )
-        try:
-            with connection.cursor() as cursor:
-                sql = (
-                    "INSERT INTO qr_chatbot (question, answer, source, created_at) "
-                    "VALUES (%s, %s, %s, %s)"
-                )
-                cursor.execute(sql, (question or "", answer or "", source or "", datetime.now()))
-        finally:
-            connection.close()
-    except Exception as e:
-        print(f"❌ MySQL save error: {e}")
-
-
-# =============================================================================
-# ADVANCED COLLECTION DECISION
-# =============================================================================
-
-def decide_collection_intelligently(query: str) -> str:
-    """
-    Decide which collection to search based on query analysis.
-    More efficient than using LLM for every decision.
-    """
-    normalized = TopicClassifier.normalize(query)
-    
-    # Property-related keywords -> biens
-    property_keywords = [
-        "appartement", "maison", "villa", "terrain", "studio", "duplex", "local",
-        "f1", "f2", "f3", "f4", "f5", "f6", "pièce", "piece", "chambre", "m2", "m²",
-        "superficie", "surface", "louer", "location", "acheter", "achat", "vente",
-        "prix", "budget", "dzd", "million", "oran", "alger", "constantine"
-    ]
-    
-    # Article-related keywords -> articles
-    article_keywords = [
-        "comment", "étape", "etape", "procédure", "procedure", "démarche", "demarche",
-        "document", "loi", "droit", "juridique", "conseil", "guide", "fiscalité",
-        "fiscalite", "réglementation", "reglementation", "investir", "investissement",
-        "quoi faire", "quels sont", "expliquer", "expliquez"
-    ]
-    
-    property_score = sum(1 for kw in property_keywords if kw in normalized)
-    article_score = sum(1 for kw in article_keywords if kw in normalized)
-    
-    if property_score > article_score:
-        return "biens"
-    elif article_score > property_score:
-        return "articles"
-    else:
-        # Default to biens for ambiguous queries
-        return "biens"
-
-
-def _extract_points_from_qdrant_response(res) -> List[Any]:
-    """Extract points from various Qdrant response formats."""
-    if not res:
-        return []
-    if hasattr(res, "points"):
-        return res.points or []
-    if isinstance(res, list):
-        return res
-    if isinstance(res, dict):
-        for key in ("result", "points", "data", "hits", "items"):
-            v = res.get(key)
-            if isinstance(v, list):
-                return v
-    return []
+        print(f"Failed to send log: {e}")
 
-
-def _get_payload_from_hit(hit) -> Dict:
-    """Extract payload from a search hit."""
-    if not hit:
-        return {}
-    if hasattr(hit, "payload"):
-        return hit.payload or {}
-    if isinstance(hit, dict):
-        return hit.get("payload", {})
-    return {}
-
-
-def retrieve_context(query: str, top_k: int = 8) -> Tuple[str, str]:
-    """
-    Retrieve relevant context from Qdrant.
-    Returns: (context_text, collection_name)
-    """
+# --- Helper Functions ---
+def retrieve_context(query, top_k=3):
     if not qdrant_client:
-        print("❌ Qdrant client not initialized.")
-        return "", ""
-    
-    query_embedding = get_embedding_via_voyage_ai(query)
-    if query_embedding is None:
-        print("❌ Failed to generate embedding.")
-        return "", ""
-    
-    chosen_collection = decide_collection_intelligently(query)
-    print(f"📚 Searching collection: {chosen_collection}")
-    
+        return ""
     try:
-        # Try query_points first
-        res = qdrant_client.query_points(
-            collection_name=chosen_collection,
-            query=query_embedding,
-            limit=top_k * 2,
-            with_payload=True,
-            with_vectors=False,
+        query_embedding = embedding_model.encode(query).tolist()
+        search_results = qdrant_client.query_points(
+            collection_name="RAG JEA",
+            query_vector=query_embedding,
+            limit=top_k,
+            with_payload=True
         )
-        points = _extract_points_from_qdrant_response(res)
-    except Exception as e:
-        print(f"⚠️ query_points failed: {e}")
-        try:
-            res = qdrant_client.search(
-                collection_name=chosen_collection,
-                query_vector=query_embedding,
-                limit=top_k * 2,
-                with_payload=True,
-            )
-            points = res if isinstance(res, list) else []
-        except Exception as e2:
-            print(f"❌ Search failed: {e2}")
-            return "", ""
-    
-    if not points:
-        return "", chosen_collection
-    
-    # Build context from results
-    context_parts = []
-    for idx, p in enumerate(points[:top_k], 1):
-        payload = _get_payload_from_hit(p)
-        
-        if chosen_collection == "biens":
-            titre = payload.get('Titre') or payload.get('titre') or ""
-            ville = payload.get('Ville') or payload.get('ville_norm') or ""
-            superficie = payload.get('superficie') or payload.get('Superficie') or ""
-            pieces = payload.get('pieces') or payload.get('Nombre de pièces') or ""
-            prix = payload.get('prix') or payload.get('Prix') or ""
-            ref = payload.get('REF') or payload.get('ref') or ""
-            url = payload.get('URL') or payload.get('url') or ""
-            commodites = payload.get('commodites') or []
-            
-            parts = [f"**{idx}. {titre}**"] if titre else [f"**{idx}. Bien immobilier**"]
-            if ville:
-                parts.append(f"📍 Ville: {ville}")
-            if superficie:
-                parts.append(f"📐 Superficie: {superficie} m²")
-            if pieces:
-                parts.append(f"🚪 Pièces: {pieces}")
-            if prix:
-                parts.append(f"💰 Prix: {prix}")
-            if commodites and isinstance(commodites, list):
-                parts.append(f"✨ Commodités: {', '.join(commodites[:5])}")
-            if ref:
-                parts.append(f"🔖 Réf: {ref}")
-            if url:
-                parts.append(f"🔗 Lien: {url}")
-            
-            context_parts.append("\n".join(parts))
-        
-        else:  # articles
-            titre = payload.get('Titre') or payload.get('titre') or ""
-            date_pub = payload.get('Date de Publication') or payload.get('date') or ""
-            contenu = payload.get('Contenu Texte') or payload.get('contenu') or ""
-            url = payload.get('URL') or payload.get('url') or ""
-            
-            parts = [f"**Article: {titre}**"] if titre else ["**Article**"]
-            if date_pub:
-                parts.append(f"📅 Date: {date_pub}")
-            if contenu:
-                # Truncate long content
-                excerpt = contenu[:600] + "..." if len(contenu) > 600 else contenu
-                parts.append(f"📝 {excerpt}")
-            if url:
-                parts.append(f"🔗 Lien: {url}")
-            
-            context_parts.append("\n".join(parts))
-    
-    context = "\n\n---\n\n".join(context_parts)
-    return context, chosen_collection
-
 
-# =============================================================================
-# LLM QUERY
-# =============================================================================
+        context_parts = []
+        for hit in search_results:
+            payload = hit.payload
+            if payload and 'text' in payload:
+                doc_text = payload['text']
+                if 'link' in payload and payload['link']:
+                    doc_text += f"\n(Source du lien : {payload['link']})"
+                context_parts.append(doc_text)
+        if context_parts:
+            context = "\n\n---\n\n".join(context_parts)
+            print(f"✅ Retrieved {len(context_parts)} relevant documents with links")
+            return context
+        else:
+            print("⚠️ No relevant documents found")
+            return ""
+    except Exception as e:
+        print(f"❌ Error retrieving context: {e}")
+        return ""
 
-def query_groq_llm(messages: List[Dict], tokens_max: int,
-                   temperature: float, top_p: float) -> str:
-    """Query Groq LLM and return complete response."""
+# --- Groq (DeepSeek) Function ---
+def query_groq_deepseek(messages, tokens_max, temperature, top_p):
     if not groq_client:
-        print("❌ Groq client not initialized")
-        return "❌ Erreur : Service LLM non disponible."
-
+        yield "❌ Erreur : Clé API Groq non configurée."
+        return
     try:
-        print(f"🤖 Calling Groq LLM with {len(messages)} messages...")
         chat_completion = groq_client.chat.completions.create(
-            model="openai/gpt-oss-safeguard-20b",
+            model="llama3-70b-8192",
             messages=messages,
             temperature=temperature,
             max_tokens=int(tokens_max),
             top_p=top_p,
             stream=True,
         )
-
         full_response = ""
         for chunk in chat_completion:
-            if chunk.choices and chunk.choices[0].delta:
-                delta = chunk.choices[0].delta.content or ""
-                full_response += delta
-
-        if not full_response.strip():
-            print("⚠️ LLM returned empty response")
-            return "Je n'ai pas pu générer une réponse. Veuillez réessayer."
-
-        print(f"✅ LLM response received: {len(full_response)} chars")
-        return full_response.strip()
+            delta = chunk.choices[0].delta.content or ""
+            full_response += delta
+            yield full_response.strip()
     except Exception as e:
-        print(f"❌ LLM Error: {type(e).__name__}: {e}")
-        return f"❌ Erreur LLM: {e}"
-
-
-# =============================================================================
-# IMPROVED SYSTEM PROMPT
-# =============================================================================
-
-SYSTEM_PROMPT = """Vous êtes un assistant immobilier expert spécialisé EXCLUSIVEMENT en immobilier en Algérie.
-
-RÈGLES STRICTES :
-1. Répondez UNIQUEMENT aux questions liées à l'immobilier en Algérie
-2. Si une question est hors sujet (cuisine, sport, politique, etc.), refusez poliment et rappelez votre spécialisation
-3. Basez vos réponses UNIQUEMENT sur le contexte fourni
-4. Si le contexte ne contient pas l'information demandée, dites clairement que vous n'avez pas cette information
-
-FORMAT DE RÉPONSE :
-- Soyez concis et direct
-- Structurez clairement les informations
-- Pour les biens : numérotez-les (1., 2., 3.)
-- Mettez en évidence les prix avec 💰
-- Incluez les liens sources quand disponibles
-- N'inventez JAMAIS de données ou de liens
-
-COMPORTEMENT :
-- Ton professionnel mais accessible
-- Pas de markdown excessif (évitez ###, ****, etc. sauf pour mettre en valeur)
-- Réponses en français
-- Pas de commentaires sur votre processus interne"""
-
-
-# =============================================================================
-# MAIN RESPONSE FUNCTION
-# =============================================================================
-
-def respond(message: str, tokens_max: int, temperature: float, top_p: float) -> Tuple[str, str]:
-    """
-    Main response function with topic filtering and intent detection.
-    """
-    message = message.strip()
-    if not message:
-        return "Veuillez saisir une question.", ""
-    
-    # Step 1: Topic Classification
-    is_on_topic, category, confidence = TopicClassifier.classify(message)
-    
-    if not is_on_topic:
-        off_topic_response = TopicClassifier.get_off_topic_response()
-        log_interaction(message, off_topic_response, "OFF_TOPIC", tokens_max, temperature, top_p)
-        try:
-            save_chat_to_mysql(message, off_topic_response, "off_topic")
-        except:
-            pass
-        return off_topic_response, "Question hors sujet détectée"
-    
-    # Step 2: Intent Detection
-    intent, intent_confidence = IntentDetector.detect(message)
-    
-    # Handle simple intents directly
-    if intent == "greeting":
-        response = ResponseTemplates.GREETING
-        log_interaction(message, response, "GREETING", tokens_max, temperature, top_p)
-        save_chat_to_mysql(message, response, "greeting")
-        return response, ""
-    
-    if intent == "thanks":
-        response = ResponseTemplates.THANKS
-        log_interaction(message, response, "THANKS", tokens_max, temperature, top_p)
-        save_chat_to_mysql(message, response, "thanks")
-        return response, ""
-    
-    if intent == "goodbye":
-        response = ResponseTemplates.GOODBYE
-        log_interaction(message, response, "GOODBYE", tokens_max, temperature, top_p)
-        save_chat_to_mysql(message, response, "goodbye")
-        return response, ""
-    
-    # Step 3: Retrieve Context
-    context, collection = retrieve_context(message, top_k=8)
-    
-    if not context:
-        response = ResponseTemplates.NO_RESULTS
-        log_interaction(message, response, "NO_CONTEXT", tokens_max, temperature, top_p)
-        save_chat_to_mysql(message, response, "no_results")
-        return response, "Aucun résultat trouvé"
-    
-    # Step 4: Generate Response with LLM
-    user_content = f"""Question de l'utilisateur : {message}
-
-Contexte pertinent de notre base de données :
+        yield f"❌ Erreur GROQ: {e}"
+
+# --- System message ---
+system_message = (
+    "Vous êtes un assistant spécialisé en immobilier algérien. "
+    "Pour chaque bien ou idée, commence une nouvelle ligne avec deux sauts de ligne (\\n\\n). "
+    "Numérote chaque bien (1., 2., 3., etc.). "
+    "Si un lien source est fourni dans le contexte (ex: Source du lien : https://...), mentionne ce lien à la fin de la description du bien. N'invente jamais de lien. "
+    "N'utilise pas de balises HTML ni de markdown, seulement du texte brut et des liens. "
+)
 
-{context}
+# --- Main Respond Function (Groq/DeepSeek used) ---
+def respond(message, tokens_max, temperature, top_p):
+    context = retrieve_context(message)
+    enhanced_system_message = system_message
+    if context:
+        enhanced_system_message += f"\n\nContexte pertinent de la base de données:\n\n{context}"
+    else:
+        enhanced_system_message += "\n\nAucun contexte spécifique trouvé dans la base de données. Fournissez des informations générales sur l'immobilier algérien."
 
-Répondez de manière précise et structurée en vous basant UNIQUEMENT sur ce contexte."""
+    messages = [{"role": "system", "content": enhanced_system_message}]
+    messages.append({"role": "user", "content": message})
 
-    messages = [
-        {"role": "system", "content": SYSTEM_PROMPT},
-        {"role": "user", "content": user_content}
-    ]
-    
-    response = query_groq_llm(messages, tokens_max, temperature, top_p)
-    
-    # Step 5: Add relevant links based on collection
-    if collection == "biens":
-        response += "\n\n---\n📋 Explorer plus de biens : https://jacheteenalgerie.com/liste-des-biens/"
-        response += "\n📅 Besoin d'accompagnement ? : https://jacheteenalgerie.com/nos-offres-immobilier/"
-    elif collection == "articles":
-        response += "\n\n---\n📅 Besoin d'accompagnement ? : https://jacheteenalgerie.com/nos-offres-immobilier/"
+    response = ""
+    for token in query_groq_deepseek(messages, tokens_max, temperature, top_p):
+        response = token
+    log_interaction(message, response, context, tokens_max, temperature, top_p)
     
-    # Step 6: Log and save
-    log_interaction(message, response, context[:500], tokens_max, temperature, top_p)
-    try:
-        save_chat_to_mysql(message, response, collection)
-    except Exception as e:
-        print(f"⚠️ MySQL save failed: {e}")
-    
-    return response, context
+    return response.strip(), context
 
-
-# =============================================================================
-# GRADIO API
-# =============================================================================
+# --- Gradio UI ---
+example_questions = [
+    "Quelles sont les étapes pour acheter un appartement en Algérie ?",
+    "Comment fonctionne la location immobilière en Algérie ?",
+    "Quels sont les droits des locataires ?",
+    "Informations sur les prix immobiliers à Oran",
+    "Documents nécessaires pour un achat immobilier"
+]
+formatted_examples = [[q, 1000, 0.7, 0.95] for q in example_questions]
 
 demo = gr.Interface(
     fn=respond,
     inputs=[
-        gr.Textbox(label="question"),
-        gr.Number(value=1024, label="tokens_max"),
-        gr.Number(value=0.3, label="temperature"),
-        gr.Number(value=0.85, label="top_p"),
-    ],
-    outputs=[
-        gr.Textbox(label="response"),
-        gr.Textbox(label="context"),
+        gr.Textbox(label="Message utilisateur"),
+        gr.Slider(minimum=1, maximum=2048, value=1000, step=1, label="Tokens maximum"),
+        gr.Slider(minimum=0.1, maximum=2.0, value=0.7, step=0.1, label="Température"),
+        gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p (nucleus sampling)"),
     ],
-    api_name="predict",
-    concurrency_limit=10  # Allow up to 10 concurrent requests
+    outputs=[gr.Textbox(label="Réponse"), gr.Textbox(label="Contexte")],
+    title="🏠 Assistant Immobilier Algérien - RAG",
+    description="""
+    **Assistant intelligent alimenté par une base de données de connaissances**
+    Posez vos questions sur l'immobilier en Algérie. L'assistant consulte une base de données 
+    de documents pertinents pour vous fournir des réponses précises et actualisées.
+    💡 **Exemples :** Achat immobilier, location, droits des locataires, prix à Oran...
+    🔧 **Système :** DeepSeek via Groq + Qdrant Cloud + all-MiniLM-L6-v2
+    """,
+    examples=formatted_examples,
+    cache_examples=False,
+    theme=gr.themes.Soft(),
 )
 
 if __name__ == "__main__":
-    demo.launch(
-        server_name="0.0.0.0",
-        server_port=7860,
-        show_error=True,
-        max_threads=20  # Enable multi-threading for concurrent requests
-    )
+    demo.launch(server_name="0.0.0.0", server_port=7860, share=True, show_error=True)

requirements.txt

@@ -13,6 +13,3 @@ qdrant-client
 groq
 gradio>=4.16.0
 google-generativeai>=0.3.2
-fastapi
-uvicorn
-pymysql