Avatar IA pour le service client : remplacer sans perdre l'humain

┌─────────────────────────────────────────────┐
│              INPUT CHANNELS                 │
│  Web Chat │ WhatsApp │ Email │ Phone        │
└──────────────────┬──────────────────────────┘
                   │
          ┌────────▼────────┐
          │   ORCHESTRATOR   │
          │   (AI Router)    │
          └────────┬────────┘
                   │
    ┌──────────────┼──────────────┐
    │              │              │
┌───▼───┐   ┌─────▼─────┐  ┌────▼────┐
│  LLM  │   │  CRM API │  │KNOWLEDGE│
│(Claude)│  │(History) │  │  BASE   │
└───┬───┘   └─────┬─────┘  └────┬────┘
    │              │              │
    └──────────────┼──────────────┘
                   │
          ┌────────▼────────┐
          │  RESPONSE        │
          │  GENERATION      │
          └────────┬────────┘
                   │
          ┌────────▼────────┐
          │  QUALITY FILTER  │
          │  + ESCALATION    │
          └────────┬────────┘
                   │
          ┌────────▼────────┐
          │  CUSTOMER        │
          │  RESPONSE        │
          └─────────────────┘

import httpx
import json

class AvatarServiceClient:
    def __init__(self, brand_config, crm_client):
        self.brand = brand_config
        self.crm = crm_client
        self.api_url = "https://openrouter.ai/api/v1/chat/completions"

    def build_system_prompt(self, customer_data):
        return f"""You are {self.brand['name']}, the virtual assistant for {self.brand['company']}.

PERSONALITY:
- Tone: {self.brand['tone']}
- Values: {', '.join(self.brand['values'])}
- Style: {self.brand['style']}

RULES:
- Always greet the customer by name
- Never invent product information
- If unsure, escalate to a human
- Acknowledge the customer's emotion before offering a solution
- Maximum 3 sentences per response unless technical explanation is needed

CURRENT CUSTOMER:
- Name: {customer_data['name']}
- Status: {customer_data['tier']}
- Last Order: {customer_data['last_order']}
- Previous Tickets: {customer_data['ticket_count']}
- Detected Sentiment: {customer_data.get('sentiment', 'neutral')}
"""

    def build_knowledge_context(self, query, top_k=5):
        # Vector search in the knowledge base
        results = self.knowledge_base.search(query, top_k=top_k)
        context = "\n\nRELEVANT KNOWLEDGE BASE:\n"
        for doc in results:
            context += f"- {doc['title']}: {doc['content']}\n"
        return context

    async def respond(self, customer_id, message, conversation_history):
        # 1. Fetch customer data
        customer = await self.crm.get_customer(customer_id)

        # 2. Analyze sentiment
        customer['sentiment'] = self.analyze_sentiment(message)

        # 3. Build the prompt
        system = self.build_system_prompt(customer)
        system += self.build_knowledge_context(message)

        # 4. Prepare messages
        messages = [{"role": "system", "content": system}]
        messages.extend(conversation_history[-10:])  # Keep last 10 exchanges
        messages.append({"role": "user", "content": message})

        # 5. LLM call
        response = await httpx.AsyncClient().post(
            self.api_url,
            headers={"Authorization": f"Bearer {self.api_key}"},
            json={
                "model": "anthropic/claude-sonnet-4-20250514",
                "messages": messages,
                "max_tokens": 500,
                "temperature": 0.3  # Low for consistency
            }
        )

        result = response.json()
        answer = result['choices'][0]['message']['content']

        # 6. Check if escalation is needed
        if self.needs_escalation(answer, customer['sentiment']):
            return self.escalate_to_human(customer_id, conversation_history)

        return answer

# Example: Luxury brand
luxury_brand = {
    "name": "Élise",
    "company": "Maison Lumière",
    "tone": "refined, warm, attentive",
    "values": ["excellence", "discretion", "bespoke"],
    "style": "formal address, elegant phrasing, no technical jargon"
}

# Example: Tech startup
tech_startup = {
    "name": "Max",
    "company": "DevFlow",
    "tone": "casual, direct, enthusiastic",
    "values": ["transparency", "speed", "innovation"],
    "style": "informal, moderate emojis, tech references allowed"
}

# Example: B2B corporate
b2b_corp = {
    "name": "Sophie",
    "company": "IndustraPro",
    "tone": "professional, factual, reassuring",
    "values": ["reliability", "expertise", "partnership"],
    "style": "formal address, data-driven, industry references"
}

CUSTOMER: "I’ve been waiting for my package for 3 WEEKS!! This is unacceptable,
           I want an IMMEDIATE refund!"

❌ Classic Chatbot:
"I understand. Can you provide your order number?"

✅ AI Avatar:
"I completely understand your frustration, Marc. Three weeks is far too long,
and I sincerely apologize for this situation. I see your order #4892—it’s been
stuck at the sorting center since February 12. Here’s what I’ll do right now:
I’ll expedite the shipment AND give you a €15 credit for your next order. If
the package doesn’t arrive by Friday, I’ll process a full refund. Does that work for you?"

CUSTOMER: "Your software crashes every time I export a PDF since the update.
           I lost 2 hours of work. This is the second time this month."

✅ AI Avatar:
"Ouch—losing 2 hours of work is infuriating, Thomas. And the fact that it’s
happening again, I totally get why that’s eroding your trust. I’ve identified
the bug—it affects PDF exports over 50 pages in v3.2.1. Our team pushed a fix
this morning (v3.2.2). Can you update and try again? In the meantime, I’ll
email you a backup of your last document. I’m also flagging your case as high
priority with the product manager."

CUSTOMER: "My e-commerce site is DOWN. It’s peak sales season—I’m losing
           thousands per minute!!"

✅ AI Avatar:
"Emergency acknowledged, Mr. Fabre. I confirm your instance has been down
since 14:32. I’m immediately escalating to our Tier 2 infrastructure team.
In parallel, I’m activating your custom maintenance page so visitors can leave
their emails. An engineer will contact you within the next 5 minutes. I’ll stay
on the line with you."

ESCALATION_RULES = {
    "triggers": [
        # Sentiment-based
        {"sentiment": "furious", "threshold": 0.9, "priority": "critical"},
        {"sentiment": "frustrated", "threshold": 0.7, "priority": "high"},

        # Topic-based
        {"keywords": ["legal", "lawyer", "sue"], "priority": "critical"},
        {"keywords": ["refund", "chargeback"], "priority": "high"},

        # Confidence-based
        {"llm_confidence": <0.6, "priority": "medium"},

        # Repetition-based
        {"repeated_issue": True, "repetitions": 3, "priority": "high"}
    ],

    "actions": {
        "critical": {
            "route_to": "senior_support",
            "timeout": 300,  # 5 minutes
            "customer_message": "I’ve escalated this to our senior team. {agent_name} will contact you within 5 minutes."
        },
        "high": {
            "route_to": "tier_2_support",
            "timeout": 900,  # 15 minutes
            "customer_message": "I’m transferring you to a specialist. {agent_name} will assist you shortly."
        }
    }
}