User asks: "How do I schedule a campaign?"
    │
    ▼
┌──────────────────────────────────────────┐
│  Step 1: Keyword Extraction               │
│  Extract key terms: "schedule", "campaign" │
└────────────────────┬─────────────────────┘
                     │
┌────────────────────▼─────────────────────┐
│  Step 2: Local Search (content.json)      │
│  Search through 50+ Q&A pairs             │
│  Score each by keyword overlap             │
│  Return top 3 matches                      │
└────────────────────┬─────────────────────┘
                     │
                     ├── Found good matches? ──► Return local answer
                     │
                     └── No good matches? ──┐
                                            │
┌───────────────────────────────────────────▼──┐
│  Step 3: AI Fallback (OpenRouter)             │
│  Send question + platform context to LLM      │
│  Model: stepfun/step-3.5-flash OR              │
│         google/gemini-2.5-flash               │
│  Get AI-generated answer                       │
└──────────────────────────────────────────────┘

{
  "platform_help": [
    {
      "question": "How do I create a campaign?",
      "keywords": ["create", "campaign", "new", "start"],
      "answer": "To create a campaign:\n1. Go to Campaigns page\n2. Click 'New Campaign'\n3. Select a template\n4. Choose your target audience by tags\n5. Set the schedule\n6. Click 'Send' or 'Schedule'"
    },
    {
      "question": "How do I import contacts?",
      "keywords": ["import", "contacts", "csv", "upload", "bulk"],
      "answer": "Go to Contacts → Import. Upload a CSV file with columns: phone, name, email. You can also add tags during import."
    }
  ]
}

def search_local_content(query):
    query_words = set(query.lower().split())
    results = []
    
    for item in content['platform_help']:
        # Count how many keywords match
        keyword_matches = len(query_words & set(item['keywords']))
        
        # Also check for phrase overlap in the question
        question_words = set(item['question'].lower().split())
        question_overlap = len(query_words & question_words)
        
        score = keyword_matches * 2 + question_overlap
        
        if score > 0:
            results.append((score, item))
    
    # Sort by score, return top 3
    results.sort(key=lambda x: x[0], reverse=True)
    return results[:3]

def get_ai_response(question, context_snippets):
    system_prompt = """You are MetaPilot Assistant, a helpful AI that assists users 
    with the MetaPilot WhatsApp marketing platform. Answer questions about:
    - Campaign management
    - Contact management  
    - Template creation
    - WhatsApp API integration
    - Analytics and reporting
    
    Keep answers concise and actionable. If you don't know, say so."""
    
    response = requests.post(
        "https://openrouter.ai/api/v1/chat/completions",
        headers={
            "Authorization": f"Bearer {OPENROUTER_API_KEY}",
            "Content-Type": "application/json"
        },
        json={
            "model": "stepfun/step-3.5-flash",
            "messages": [
                {"role": "system", "content": system_prompt},
                {"role": "user", "content": f"Context: {context_snippets}\n\nQuestion: {question}"}
            ],
            "max_tokens": 500,
            "temperature": 0.3  # Low temperature for factual answers
        }
    )
    
    return response.json()['choices'][0]['message']['content']

POST /api/chatbot/ask/

Request:
{
    "question": "How do I schedule a campaign for next Monday?"
}

Response:
{
    "answer": "To schedule a campaign:\n1. Go to Campaigns...",
    "source": "local",  // or "ai"
    "confidence": 0.85
}

Customer sends WhatsApp message
    │
    ▼
Webhook receives message
    │
    ▼
WA_CHATBOT_ENABLED = True?
    │
    ├── No → Skip (message stored but no auto-reply)
    │
    └── Yes ─┐
             │
┌────────────▼──────────────────────────────────┐
│  WAChatbotService.generate_reply()             │
│                                                │
│  1. Load conversation history (last 10 msgs)   │
│  2. Build system prompt with business context   │
│  3. Select AI model based on message type:      │
│     • Text → stepfun/step-3.5-flash            │
│     • Image → openai/gpt-4o (vision)           │
│  4. Call OpenRouter API                          │
│  5. Return generated reply                       │
└────────────────────┬─────────────────────────┘
                     │
┌────────────────────▼──────────────────────────┐
│  Send reply via WhatsApp API                    │
│  WhatsAppService.send_text(to, reply)           │
└───────────────────────────────────────────────┘

def build_messages(self, tenant, customer_phone, new_message):
    # Load last 10 messages from this conversation
    recent_messages = InboxMessage.objects.filter(
        conversation__tenant=tenant,
        conversation__customer_phone=customer_phone
    ).order_by('-created_at')[:10]
    
    messages = [
        {
            "role": "system",
            "content": self.get_system_prompt(tenant)
        }
    ]
    
    # Add conversation history
    for msg in reversed(recent_messages):
        role = "assistant" if msg.direction == 'OUTBOUND' else "user"
        messages.append({
            "role": role,
            "content": msg.content_json.get('text', '')
        })
    
    # Add the new message
    messages.append({
        "role": "user",
        "content": new_message
    })
    
    return messages

def get_system_prompt(self, tenant):
    return f"""You are a customer service assistant for {tenant.name}.
    Business type: {tenant.business_type}
    
    Guidelines:
    - Be friendly and professional
    - Answer questions about products and services
    - If you can't help, suggest contacting support
    - Never share customer data
    - Keep responses under 200 words
    - Use the customer's language (auto-detect)
    """

def process_image_message(self, image_url, caption=""):
    response = requests.post(
        "https://openrouter.ai/api/v1/chat/completions",
        json={
            "model": "openai/gpt-4o",  # Vision-capable model
            "messages": [
                {
                    "role": "user",
                    "content": [
                        {
                            "type": "text",
                            "text": f"A customer sent this image. {caption}. Describe what you see and respond helpfully."
                        },
                        {
                            "type": "image_url",
                            "image_url": {"url": image_url}
                        }
                    ]
                }
            ],
            "max_tokens": 300
        }
    )
    return response.json()['choices'][0]['message']['content']

POST /api/wa-chatbot/webhook/{tenant_id}/

# Required for AI features
OPENROUTER_API_KEY=sk-or-v1-xxx

# Feature flags
WA_CHATBOT_ENABLED=True      # Enable WhatsApp auto-replies
CHATBOT_MAX_TOKENS=500        # Max response length
CHATBOT_TEMPERATURE=0.3       # Creativity (0=factual, 1=creative)