IA conversationnelle : meilleures pratiques pour le développement de chatbots

Les systèmes d’IA conversationnelle permettent une interaction homme-machine via le langage naturel. Dans ce guide, nous examinons les stratégies de développement de chatbots pour les entreprises.

Architecture du chatbot

Composants principaux

1┌────────────────────────────────────────────────────┐
2│                    User Interface                   │
3└───────────────────────┬────────────────────────────┘
4                        │
5┌───────────────────────▼────────────────────────────┐
6│              Input Processing Layer                 │
7│  ┌──────────┐  ┌──────────┐  ┌────────────────┐   │
8│  │  Speech  │  │   Text   │  │   Multimodal   │   │
9│  │  to Text │  │ Normalize│  │   Processing   │   │
10│  └──────────┘  └──────────┘  └────────────────┘   │
1130: └───────────────────────┬────────────────────────────┘
12                        │
13┌───────────────────────▼────────────────────────────┐
14│                 Dialog Manager                      │
15│  ┌──────────┐  ┌──────────┐  ┌────────────────┐   │
16│  │  Intent  │  │  State   │  │    Context     │   │
17│  │ Detection│  │ Manager  │  │    Handler     │   │
18│  └──────────┘  └──────────┘  └────────────────┘   │
1938: └───────────────────────┬────────────────────────────┘
20                        │
21┌───────────────────────▼────────────────────────────┐
22│              Response Generation                    │
23│  ┌──────────┐  ┌──────────┐  ┌────────────────┐   │
24│  │   LLM    │  │ Template │  │     Action     │   │
25│  │ Response │  │  Engine  │  │    Executor    │   │
26│  └──────────┘  └──────────┘  └────────────────┘   │
2746: └────────────────────────────────────────────────────┘

Gestion du contexte

Mémoire de conversation

1from typing import List, Dict
2from datetime import datetime
3
4class ConversationMemory:
5    def __init__(self, max_turns: int = 10):
6        self.max_turns = max_turns
7        self.history: List[Dict] = []
8        self.metadata: Dict = {}
9    
10    def add_message(self, role: str, content: str):
11        message = {
12            "role": role,
13            "content": content,
14            "timestamp": datetime.now().isoformat()
15        }
16        self.history.append(message)
17        
18        # Clean old messages
19        if len(self.history) > self.max_turns * 2:
20            self.history = self.history[-self.max_turns * 2:]
21    
22    def get_context(self) -> List[Dict]:
23        return [{"role": m["role"], "content": m["content"]} 
24                for m in self.history]
25    
26    def summarize(self, llm_client) -> str:
27        """Summarize long conversations"""
28        if len(self.history) < 10:
29            return None
30        
31        response = llm_client.chat.completions.create(
32            model="gpt-4-turbo",
33            messages=[
34                {"role": "system", "content": "Summarize this conversation."},
35                {"role": "user", "content": str(self.history)}
36            ]
37        )
38        return response.choices[0].message.content

État de session

1class SessionState:
2    def __init__(self, session_id: str):
3        self.session_id = session_id
4        self.user_info: Dict = {}
5        self.collected_data: Dict = {}
6        self.current_intent: str = None
7        self.current_step: str = None
8        self.awaiting_input: str = None
9    
10    def set_slot(self, name: str, value: any):
11        self.collected_data[name] = value
12    
13    def get_slot(self, name: str) -> any:
14        return self.collected_data.get(name)
15    
16    def clear_slots(self):
17        self.collected_data = {}
18    
19    def is_complete(self, required_slots: List[str]) -> bool:
20        return all(slot in self.collected_data for slot in required_slots)
21## Classification d’intention
22
23### Détection d’intention basée sur les LLM
24
25```python
26def detect_intent(user_message: str, available_intents: List[Dict]) -> Dict:
27    intent_descriptions = "\n".join([
28        f"- {intent['name']}: {intent['description']}"
29        for intent in available_intents
30    ])
31    
32    response = client.chat.completions.create(
33        model="gpt-4-turbo",
34        response_format={"type": "json_object"},
35        messages=[
36            {
37                "role": "system",
38                "content": f"""Determine the intent of the user message.
39                
40Available intents:
41{intent_descriptions}
42
43Respond in JSON format:
44{{"intent": "intent_name", "confidence": 0.0-1.0, "entities": {{}}}}
45"""
46            },
47            {"role": "user", "content": user_message}
48        ]
49    )
50    
51    return json.loads(response.choices[0].message.content)

Détection d’intention hybride

1from sklearn.feature_extraction.text import TfidfVectorizer
2from sklearn.linear_model import LogisticRegression
3
4class HybridIntentClassifier:
5    def __init__(self):
6        self.ml_classifier = None
7        self.vectorizer = TfidfVectorizer()
8    
9    def train_ml(self, texts: List[str], labels: List[str]):
10        X = self.vectorizer.fit_transform(texts)
11        self.ml_classifier = LogisticRegression()
12        self.ml_classifier.fit(X, labels)
13    
14    def classify(self, text: str) -> Dict:
15        # ML prediction
16        X = self.vectorizer.transform([text])
17        ml_intent = self.ml_classifier.predict(X)[0]
18        ml_confidence = max(self.ml_classifier.predict_proba(X)[0])
19        
20        # Ask LLM if confidence is low
21        if ml_confidence < 0.7:
22            llm_result = detect_intent(text, self.available_intents)
23            return llm_result
24        
25        return {"intent": ml_intent, "confidence": ml_confidence}

Flux de dialogue

Machine à états

1from enum import Enum
2
3class OrderState(Enum):
4    START = "start"
5    PRODUCT_SELECTION = "product_selection"
6    QUANTITY = "quantity"
7    ADDRESS = "address"
8    CONFIRMATION = "confirmation"
9    COMPLETE = "complete"
10
11class OrderFlow:
12    def __init__(self):
13        self.state = OrderState.START
14        self.data = {}
15    
16    def process(self, user_input: str, intent: str) -> str:
17        if self.state == OrderState.START:
18            if intent == "order":
19                self.state = OrderState.PRODUCT_SELECTION
20                return "Quel produit souhaitez-vous commander ?"
21        
22        elif self.state == OrderState.PRODUCT_SELECTION:
23            self.data["product"] = user_input
24            self.state = OrderState.QUANTITY
25            return f"Vous avez choisi {user_input}. Quelle quantité souhaitez-vous ?"
26        
27        elif self.state == OrderState.QUANTITY:
28            self.data["quantity"] = int(user_input)
29            self.state = OrderState.ADDRESS
30            return "Pourriez-vous fournir votre adresse de livraison ?"
31        
32        elif self.state == OrderState.ADDRESS:
33            self.data["address"] = user_input
34            self.state = OrderState.CONFIRMATION
35            return f"Récapitulatif de la commande:\n{self.data}\nConfirmez-vous ?"
36        
37        elif self.state == OrderState.CONFIRMATION:
38            if "yes" in user_input.lower():
39                self.state = OrderState.COMPLETE
40                return "Commande reçue. Merci !"
41            else:
42                self.state = OrderState.START
43                return "Commande annulée."
44## Génération de réponses
45
46### Modèle hybride Template + LLM
47
48```python
49class ResponseGenerator:
50    def __init__(self):
51        self.templates = {
52            "greeting": [
53                "Hello! How can I help you?",
54                "Welcome! What would you like to do today?"
55            ],
56            "farewell": [
57                "Have a good day! See you again.",
58                "I'm here if you have any other questions. Goodbye!"
59            ]
60        }
61    
62    def generate(self, intent: str, context: Dict) -> str:
63        # Template for simple intents
64        if intent in self.templates:
65            return random.choice(self.templates[intent])
66        
67        # LLM for complex intents
68        return self.llm_generate(intent, context)
69    
70    def llm_generate(self, intent: str, context: Dict) -> str:
71        system_prompt = f"""You are a helpful customer service assistant.
72User intent: {intent}
73Current context: {json.dumps(context, ensure_ascii=False)}
74
75Rules:
76- Give short and concise answers
77- Be friendly but professional
78- Ask follow-up questions if necessary
79"""
80        
81        response = client.chat.completions.create(
82            model="gpt-4-turbo",
83            messages=[
84                {"role": "system", "content": system_prompt},
85                *context.get("history", [])
86            ]
87        )
88        
89        return response.choices[0].message.content

Gestion des solutions de repli

1class FallbackHandler:
2    def __init__(self):
3        self.fallback_count = 0
4        self.max_fallbacks = 3
5    
6    def handle(self, user_input: str, context: Dict) -> str:
7        self.fallback_count += 1
8        
9        if self.fallback_count >= self.max_fallbacks:
10            self.fallback_count = 0
11            return self.human_handoff(context)
12        
13        responses = [
14            "I didn't understand. Could you rephrase that?",
15            "I cannot help with this. You can ask a different question.",
16            "I think I misunderstood you. could you try again?"
17        ]
18        
19        return responses[self.fallback_count - 1]
20    
21    def human_handoff(self, context: Dict) -> str:
22        # Hand off to human agent
23        create_support_ticket(context)
24        return "I am connecting you to a customer representative. Please wait."

Analytique et surveillance

1class ChatAnalytics:
2    def __init__(self):
3        self.metrics = {
4            "total_conversations": 0,
5            "successful_resolutions": 0,
6            "fallback_rate": 0,
7            "avg_turns_per_conversation": 0,
8            "intent_distribution": {}
9        }
10    
11    def log_conversation(self, conversation: Dict):
12        self.metrics["total_conversations"] += 1
13        
14        # Intent tracking
15        for message in conversation["messages"]:
16            if "intent" in message:
17                intent = message["intent"]
18                self.metrics["intent_distribution"][intent] = \
19                    self.metrics["intent_distribution"].get(intent, 0) + 1
20        
21        # Resolution tracking
22        if conversation.get("resolved"):
23            self.metrics["successful_resolutions"] += 1
24    
25    def get_resolution_rate(self) -> float:
26        if self.metrics["total_conversations"] == 0:
27            return 0
28        return self.metrics["successful_resolutions"] / self.metrics["total_conversations"]

Checklist de mise en production

Sécurité

• Assainissement des entrées
• Limitation du débit
• Masquage des PII
• Protection contre les injections

Performance

• Temps de réponse < 2s
• Support des utilisateurs simultanés
• Couche de cache
• Auto-scalage

Expérience utilisateur

• Indicateurs de saisie
• Messages d'erreur
• Historique de conversation
• Prise en charge multilingue

Conclusion

L’IA conversationnelle est une technologie puissante qui transforme les interactions avec les clients. Vous pouvez créer des chatbots performants grâce à un flux de dialogue bien conçu, une gestion robuste des erreurs et une amélioration continue.

Chez Veni AI, nous développons des solutions de chatbot pour les entreprises.