import pandas as pd
from datasets import load_dataset
from sentence_transformers import SentenceTransformer, util
from huggingface_hub import InferenceClient
import gradio as gr
import os
import torch

# ==========================
# LOAD DATASET
# ==========================
print("1. Ingesting Data...")

try:
    raw_data = load_dataset("Kaludi/Customer-Support-Responses")
    df = pd.DataFrame(raw_data["train"])

    df = df[["query", "response"]]
    df.columns = ["customer_query", "historical_resolution"]

    print("Cloud dataset loaded successfully.")

except Exception as e:
    print(f"Cloud dataset unavailable ({e}). Using backup dataset...")

    backup_data = [
        {
            "customer_query": "How do I reset my password?",
            "historical_resolution":
            "To reset your password, navigate to login and click Forgot Password. A reset email will be sent."
        },
        {
            "customer_query": "Where is my invoice?",
            "historical_resolution":
            "Invoices are available in Billing & Payments under your account dashboard."
        },
        {
            "customer_query": "How do I cancel subscription?",
            "historical_resolution":
            "Navigate to Settings > Subscription > Cancel Plan."
        },
        {
            "customer_query": "Can I upgrade my account?",
            "historical_resolution":
            "Yes, go to Account Settings > Upgrade Plan."
        },
        {
            "customer_query": "App crashes on startup",
            "historical_resolution":
            "Please update the app and clear cache. If issue persists, contact support."
        }
    ]

    df = pd.DataFrame(backup_data)

# ==========================
# LOAD EMBEDDING MODEL
# ==========================
print("2. Loading Embedding Model...")

embedding_model = SentenceTransformer(
    "paraphrase-multilingual-MiniLM-L12-v2"
)

sample_resolutions = df["historical_resolution"].head(100).tolist()

resolution_embeddings = embedding_model.encode(
    sample_resolutions,
    convert_to_tensor=True
)

# ==========================
# CONNECT TO LLM
# ==========================
print("3. Connecting to Qwen 2.5...")

hf_token = os.environ.get("HF_TOKEN")

client = InferenceClient(
    model="Qwen/Qwen2.5-7B-Instruct",
    token=hf_token
)

# ==========================
# MAIN FUNCTION
# ==========================
def resolve_ticket(customer_query):

    if not customer_query.strip():
        return (
            "⚠️ Please enter a support ticket.",
            "",
            "0%"
        )

    query_embedding = embedding_model.encode(
        customer_query,
        convert_to_tensor=True
    )

    hits = util.semantic_search(
        query_embedding,
        resolution_embeddings,
        top_k=1
    )

    best_match = hits[0][0]
    best_match_index = best_match["corpus_id"]
    similarity_score = float(best_match["score"])

    confidence = round(similarity_score * 100, 2)

    retrieved_context = (
        df["historical_resolution"]
        .iloc[best_match_index]
    )

    messages = [
        {
            "role": "system",
            "content":
            """
            You are a professional multilingual
            customer support assistant.

            Rules:
            - Respond ONLY using the provided internal document
            - Use professional tone
            - Reply in the EXACT SAME LANGUAGE
              as the user's query
            - Be concise but helpful
            """
        },
        {
            "role": "user",
            "content":
            f"""
            Customer Query:
            {customer_query}

            Internal Document:
            {retrieved_context}
            """
        }
    ]

    try:
        response = client.chat.completions.create(
            messages=messages,
            max_tokens=250,
            temperature=0.3
        )

        final_response = (
            response.choices[0]
            .message.content.strip()
        )

        return (
            final_response,
            retrieved_context,
            f"{confidence}%"
        )

    except Exception as e:
        return (
            f"❌ Error: {str(e)}",
            retrieved_context,
            f"{confidence}%"
        )


# ==========================
# CUSTOM CSS
# ==========================
custom_css = """
body {
    background: linear-gradient(
        135deg,
        #050816,
        #0f172a
    );
}

.gradio-container {
    background:
    linear-gradient(
        135deg,
        #050816,
        #0f172a
    ) !important;

    font-family:
    'Segoe UI',
    sans-serif !important;
}

.glass-card {
    background:
    rgba(255,255,255,0.08);

    border:
    1px solid rgba(255,255,255,0.15);

    backdrop-filter:
    blur(18px);

    border-radius: 20px;
    padding: 20px;

    box-shadow:
    0 0 25px rgba(0,255,255,0.18);
}

h1 {
    text-align: center;
    color: #00f5ff;
    text-shadow:
    0px 0px 20px #00f5ff;
}

textarea {
    border-radius: 15px !important;
    border: 1px solid #00f5ff !important;

    box-shadow:
    0px 0px 15px rgba(
        0,
        245,
        255,
        0.4
    ) !important;
}

button {
    background:
    linear-gradient(
        90deg,
        #7c3aed,
        #00f5ff
    ) !important;

    color: white !important;

    border: none !important;
    border-radius: 14px !important;

    font-size: 18px !important;
    font-weight: bold !important;

    height: 55px !important;

    box-shadow:
    0 0 20px rgba(
        0,
        245,
        255,
        0.7
    ) !important;
}

button:hover {
    transform: scale(1.03);
    transition: 0.2s ease;
}

footer {
    visibility: hidden;
}
"""

# ==========================
# UI
# ==========================
with gr.Blocks(
    theme=gr.themes.Glass(),
    css=custom_css,
    title="Multilingual AI Support"
) as demo:

    gr.Markdown("""
    # 🌌 Multilingual AI Support Triage

    ### ⚡ Enterprise AI Ticket Resolution System

    🟢 **Status:** Online  
    🌍 **Languages:** 100+  
    🧠 **AI Model:** Qwen 2.5 7B  
    🔍 **Pipeline:** Retrieval-Augmented Generation (RAG)

    ---
    """)

    with gr.Row():

        with gr.Column(scale=1):

            gr.Markdown("## 📨 Customer Ticket")

            input_query = gr.Textbox(
                label="Enter support query",
                placeholder=
                "Try Spanish, Hindi, German...",
                lines=7
            )

            submit_btn = gr.Button(
                "🚀 Resolve Ticket"
            )

            gr.Markdown("### 🌍 Example Queries")

            gr.Examples(
                examples=[
                    ["¿Cómo restablezco mi contraseña?"],
                    ["Mein Konto wurde doppelt belastet"],
                    ["मेरा पासवर्ड रीसेट नहीं हो रहा"],
                    ["Comment annuler mon abonnement ?"]
                ],
                inputs=input_query
            )

        with gr.Column(scale=1):

            gr.Markdown("## 🤖 AI Resolution")

            output_reply = gr.Textbox(
                label="Generated Response",
                lines=8
            )

            retrieved_doc = gr.Textbox(
                label=
                "Retrieved Internal Document",
                lines=5
            )

            confidence_box = gr.Textbox(
                label="🎯 Match Confidence"
            )

    submit_btn.click(
        fn=resolve_ticket,
        inputs=input_query,
        outputs=[
            output_reply,
            retrieved_doc,
            confidence_box
        ],
        show_progress=True
    )

demo.launch()