utils/triage_engine.py

"""
Triage engine — calls OpenAI to replicate the n8n AI Agent logic.
In production, the Streamlit app posts to the n8n webhook; this module
runs locally for the interactive demo and HuggingFace Space.
"""
import json
import random
from datetime import datetime, timedelta

# ── Mock order database ────────────────────────────────────────────────────────
MOCK_ORDERS = {
    "ORD-789": {
        "product": "Premium Wireless Headphones",
        "status": "delayed",
        "amount": 149.99,
        "eta": (datetime.now() + timedelta(days=5)).strftime("%Y-%m-%d"),
        "carrier": "FedEx",
        "tracking": "FX123456789",
        "delay_reason": "weather disruption at hub",
    },
    "ORD-456": {
        "product": "Smart Watch Series 3",
        "status": "in_transit",
        "amount": 299.00,
        "eta": (datetime.now() + timedelta(days=2)).strftime("%Y-%m-%d"),
        "carrier": "UPS",
        "tracking": "1Z999AA10123456784",
        "delay_reason": None,
    },
    "ORD-321": {
        "product": "Laptop Stand Pro",
        "status": "delivered",
        "amount": 49.99,
        "eta": "Delivered",
        "carrier": "USPS",
        "tracking": "9400111899223974657",
        "delay_reason": None,
    },
    "ORD-654": {
        "product": "Mechanical Keyboard RGB",
        "status": "processing",
        "amount": 89.99,
        "eta": (datetime.now() + timedelta(days=7)).strftime("%Y-%m-%d"),
        "carrier": "DHL",
        "tracking": "Pending",
        "delay_reason": None,
    },
    "ORD-987": {
        "product": "Gaming Mouse X500",
        "status": "delayed",
        "amount": 79.99,
        "eta": (datetime.now() + timedelta(days=3)).strftime("%Y-%m-%d"),
        "carrier": "FedEx",
        "tracking": "FX987654321",
        "delay_reason": "high volume at sorting facility",
    },
}

SUPPORTED_LANGUAGES = [
    "Spanish", "French", "German", "Japanese", "Portuguese", "Italian",
    "Dutch", "Korean", "Chinese", "Arabic", "Hindi", "Russian", "Turkish",
    "Polish", "Swedish", "Danish", "Norwegian", "Finnish", "Greek", "Hebrew",
    "Thai", "Vietnamese", "Indonesian", "Malay", "Czech", "Hungarian",
    "Romanian", "Bulgarian", "Croatian", "Slovak",
]

# ── Demo output (no API key) ───────────────────────────────────────────────────
DEMO_OUTPUTS = {
    "TKT-001": {
        "detected_language": "Spanish",
        "language_code": "es",
        "order_id": "ORD-789",
        "order_status": "delayed — weather disruption at hub, ETA 2025-01-28",
        "sentiment": "very_angry",
        "sentiment_score": 0.94,
        "native_reply": (
            "Estimado Carlos García,\n\n"
            "Le pedimos disculpas sinceramente por los inconvenientes causados con su pedido #ORD-789. "
            "Entendemos completamente su frustración y lamentamos profundamente este retraso inaceptable.\n\n"
            "Hemos verificado que su pedido se encuentra en tránsito pero ha experimentado un retraso "
            "debido a interrupciones climáticas en nuestro centro logístico. La nueva fecha estimada "
            "de entrega es el 28 de enero de 2025.\n\n"
            "Como compensación por este inconveniente, le ofrecemos:\n"
            "• Envío gratuito en su próximo pedido\n"
            "• Descuento del 15% aplicado automáticamente\n\n"
            "Nuestro equipo de servicio al cliente de habla hispana se pondrá en contacto con usted "
            "en las próximas 2 horas para resolver esto personalmente.\n\n"
            "Atentamente,\nEquipo de Soporte Global"
        ),
        "suggested_status": "open",
        "english_summary": "Customer Carlos García is very angry about a 3-week delay on order ORD-789 (Premium Wireless Headphones). Order is delayed due to weather disruption. Customer demands immediate refund. Escalation recommended.",
    },
}

# ── OpenAI triage ──────────────────────────────────────────────────────────────
def _call_openai(ticket_id, subject, body, name, email, order_info, api_key):
    try:
        from openai import OpenAI
    except ImportError:
        return None

    client = OpenAI(api_key=api_key)

    order_context = ""
    if order_info:
        order_context = f"\n\nOrder Status Retrieved:\n{json.dumps(order_info, indent=2)}"

    schema = {
        "detected_language": "string — full language name e.g. Spanish",
        "language_code": "string — ISO 639-1 code e.g. es",
        "order_id": "string — extracted order ID or empty string",
        "order_status": "string — human readable status from order data, or 'Not found'",
        "sentiment": "string — one of: neutral, mild_concern, upset, very_angry",
        "sentiment_score": "number — 0.0 to 1.0",
        "native_reply": "string — complete professional reply in the customer's detected language",
        "suggested_status": "string — zendesk status: solved or open",
        "english_summary": "string — 1-2 sentence summary in English for the support team",
    }

    prompt = f"""You are an elite multi-lingual customer support AI for a global e-commerce company.

Ticket ID: {ticket_id}
Subject: {subject}
Customer: {name} ({email})
Message: {body}{order_context}

Analyze this ticket and return ONLY a valid JSON object matching this schema:
{json.dumps(schema, indent=2)}

Rules:
- detected_language: Full name of the language the customer wrote in
- sentiment must be exactly: neutral, mild_concern, upset, or very_angry
- native_reply must be written entirely in the customer's detected language, be empathetic, professional, and reference real order details if available
- If order is delayed or missing, acknowledge it and offer compensation
- english_summary is for the internal English-speaking team
"""

    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": prompt}],
        response_format={"type": "json_object"},
        temperature=0.3,
    )

    return json.loads(response.choices[0].message.content)


def _extract_order_id(text: str) -> str:
    import re
    patterns = [
        r'#?(ORD[-_]?\d+)',
        r'order\s+(?:number|#|id)?[\s:]*([A-Z0-9][-A-Z0-9]{3,})',
        r'\b([A-Z]{2,4}-\d{3,})\b',
    ]
    for p in patterns:
        m = re.search(p, text, re.IGNORECASE)
        if m:
            raw = m.group(1).upper().replace('_', '-')
            if not raw.startswith('ORD-'):
                raw = 'ORD-' + raw.lstrip('ORD').lstrip('-')
            return raw
    return ""


def _demo_result(ticket_id: str, body: str) -> dict:
    if ticket_id in DEMO_OUTPUTS:
        return DEMO_OUTPUTS[ticket_id]

    order_id = _extract_order_id(body)
    order_info = MOCK_ORDERS.get(order_id, {})

    # Language heuristics
    lang_map = [
        (["hola", "gracias", "pedido", "enojado", "quiero", "reembolso", "semanas"], "Spanish", "es"),
        (["bonjour", "commande", "merci", "livraison"], "French", "fr"),
        (["bestellung", "bitte", "danke", "lieferung", "enttäuschend"], "German", "de"),
        (["注文", "配送", "届いて", "確認"], "Japanese", "ja"),
        (["pedido", "chegou", "urgente", "obrigado"], "Portuguese", "pt"),
        (["ordine", "grazie", "consegna", "spedizione"], "Italian", "it"),
    ]

    body_lower = body.lower()
    detected_language, language_code = "English", "en"
    for words, lang, code in lang_map:
        if any(w in body_lower for w in words):
            detected_language, language_code = lang, code
            break

    # Sentiment heuristics
    angry_words = ["enojado", "angry", "furious", "inaceptable", "reembolso", "wut", "angry", "terrible", "horrible", "enttäuschend", "sofort"]
    upset_words = ["disappointed", "unhappy", "not happy", "upset", "still waiting", "urgente", "please resolve"]
    mild_words  = ["wondering", "checking", "voudrais", "möchte", "confirmar", "savoir"]

    if any(w in body_lower for w in angry_words) and ("!" in body or "NOW" in body.upper() or "JETZT" in body.upper()):
        sentiment = "very_angry"
        score = round(random.uniform(0.85, 0.97), 2)
    elif any(w in body_lower for w in angry_words):
        sentiment = "upset"
        score = round(random.uniform(0.65, 0.84), 2)
    elif any(w in body_lower for w in upset_words):
        sentiment = "upset"
        score = round(random.uniform(0.55, 0.75), 2)
    elif any(w in body_lower for w in mild_words):
        sentiment = "mild_concern"
        score = round(random.uniform(0.3, 0.54), 2)
    else:
        sentiment = "neutral"
        score = round(random.uniform(0.05, 0.29), 2)

    order_status = "Not found"
    if order_info:
        order_status = f"{order_info['status'].replace('_',' ').title()}"
        if order_info.get("delay_reason"):
            order_status += f" — {order_info['delay_reason']}, ETA {order_info['eta']}"
        else:
            order_status += f", ETA {order_info['eta']}"

    native_replies = {
        "es": f"Estimado/a cliente,\n\nLamentamos los inconvenientes con su pedido{' ' + order_id if order_id else ''}. Estado actual: {order_status}.\nNuestro equipo se pondrá en contacto con usted pronto.\n\nAtentamente,\nSoporte Global",
        "fr": f"Cher(e) client(e),\n\nNous sommes désolés pour les désagréments concernant votre commande{' ' + order_id if order_id else ''}. Statut actuel: {order_status}.\nNotre équipe vous contactera très prochainement.\n\nCordialement,\nSupport Global",
        "de": f"Liebe/r Kunde/Kundin,\n\nWir entschuldigen uns für die Unannehmlichkeiten mit Ihrer Bestellung{' ' + order_id if order_id else ''}. Aktueller Status: {order_status}.\nUnser Team wird sich bald bei Ihnen melden.\n\nMit freundlichen Grüßen,\nGlobaler Support",
        "ja": f"お客様へ,\n\nご注文{(' ' + order_id) if order_id else ''}に関するご不便をおかけして申し訳ございません。現在の状況: {order_status}。\n担当者よりご連絡いたします。\n\nよろしくお願いいたします,\nグローバルサポート",
        "pt": f"Caro(a) cliente,\n\nLamentamos os inconvenientes com o seu pedido{' ' + order_id if order_id else ''}. Status atual: {order_status}.\nNossa equipe entrará em contato em breve.\n\nAtenciosamente,\nSuporteGlobal",
        "en": f"Dear Customer,\n\nWe sincerely apologize for any inconvenience with your order{' ' + order_id if order_id else ''}. Current status: {order_status}.\nOur team will reach out to you shortly.\n\nBest regards,\nGlobal Support",
    }

    native_reply = native_replies.get(language_code, native_replies["en"])

    summaries = {
        "very_angry": f"Customer is very angry about order {order_id or 'N/A'} ({order_status}). Immediate attention required.",
        "upset": f"Customer is upset about order {order_id or 'N/A'} ({order_status}). Needs prompt resolution.",
        "mild_concern": f"Customer has mild concerns about order {order_id or 'N/A'} ({order_status}). Standard follow-up.",
        "neutral": f"Customer is checking on order {order_id or 'N/A'} ({order_status}). Routine inquiry.",
    }

    return {
        "detected_language": detected_language,
        "language_code": language_code,
        "order_id": order_id,
        "order_status": order_status,
        "sentiment": sentiment,
        "sentiment_score": score,
        "native_reply": native_reply,
        "suggested_status": "open" if sentiment == "very_angry" else "solved",
        "english_summary": summaries[sentiment],
    }


def run_triage(ticket_id: str, subject: str, body: str, name: str, email: str,
               api_key: str = None, demo_mode: bool = False) -> dict:
    order_id = _extract_order_id(body)
    order_info = MOCK_ORDERS.get(order_id)

    if demo_mode or not api_key:
        result = _demo_result(ticket_id, body)
    else:
        result = _call_openai(ticket_id, subject, body, name, email, order_info, api_key)
        if result is None:
            result = _demo_result(ticket_id, body)

    if order_info and not result.get("order_status"):
        result["order_status"] = order_info["status"]

    return result