import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import re
from tokenizers import normalizers # For isinstance check
from tokenizers.normalizers import Sequence, Replace, Strip
from tokenizers import Regex
import os


# --- Model & Tokenizer Configuration ---
model1_path = "https://huggingface.co/spaces/SzegedAI/AI_Detector/resolve/main/modernbert.bin"
model2_path = "https://huggingface.co/mihalykiss/modernbert_2/resolve/main/Model_groups_3class_seed12"
model3_path = "https://huggingface.co/mihalykiss/modernbert_2/resolve/main/Model_groups_3class_seed22"

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"Using device: {device}")

try:
    tokenizer = AutoTokenizer.from_pretrained("answerdotai/ModernBERT-base")

    model_1 = AutoModelForSequenceClassification.from_pretrained("answerdotai/ModernBERT-base", num_labels=41)
    model_1.load_state_dict(torch.hub.load_state_dict_from_url(model1_path, map_location=device, progress=True))
    model_1.to(device).eval()

    model_2 = AutoModelForSequenceClassification.from_pretrained("answerdotai/ModernBERT-base", num_labels=41)
    model_2.load_state_dict(torch.hub.load_state_dict_from_url(model2_path, map_location=device, progress=True))
    model_2.to(device).eval()

    model_3 = AutoModelForSequenceClassification.from_pretrained("answerdotai/ModernBERT-base", num_labels=41)
    model_3.load_state_dict(torch.hub.load_state_dict_from_url(model3_path, map_location=device, progress=True))
    model_3.to(device).eval()

except Exception as e:
    print(f"Error during model loading: {e}")
    tokenizer = None
    model_1, model_2, model_3 = None, None, None


label_mapping = {
    0: '13B', 1: '30B', 2: '65B', 3: '7B', 4: 'GLM130B', 5: 'bloom_7b',
    6: 'bloomz', 7: 'cohere', 8: 'davinci', 9: 'dolly', 10: 'dolly-v2-12b',
    11: 'flan_t5_base', 12: 'flan_t5_large', 13: 'flan_t5_small',
    14: 'flan_t5_xl', 15: 'flan_t5_xxl', 16: 'gemma-7b-it', 17: 'gemma2-9b-it',
    18: 'gpt-3.5-turbo', 19: 'gpt-35', 20: 'gpt4', 21: 'gpt4o',
    22: 'gpt_j', 23: 'gpt_neox', 24: 'human', 25: 'llama3-70b', 26: 'llama3-8b',
    27: 'mixtral-8x7b', 28: 'opt_1.3b', 29: 'opt_125m', 30: 'opt_13b',
    31: 'opt_2.7b', 32: 'opt_30b', 33: 'opt_350m', 34: 'opt_6.7b',
    35: 'opt_iml_30b', 36: 'opt_iml_max_1.3b', 37: 't0_11b', 38: 't0_3b',
    39: 'text-davinci-002', 40: 'text-davinci-003'
}

def clean_text(text: str) -> str:
    text = re.sub(r'\s{2,}', ' ', text)
    text = re.sub(r'\s+([,.;:?!])', r'\1', text)
    return text

if tokenizer:
    custom_normalizers_to_add = [
        Replace(Regex(r'(\w+)[--]\s*\n\s*(\w+)'), r"\1\2"),
        Replace(Regex(r'\s*\n\s*'), " "),
        Strip()
    ]
    current_backend_normalizer = tokenizer.backend_tokenizer.normalizer

    if current_backend_normalizer is None:
        tokenizer.backend_tokenizer.normalizer = Sequence(custom_normalizers_to_add)
    elif isinstance(current_backend_normalizer, normalizers.Sequence):
        # Extend the existing list of normalizers within the Sequence object
        current_backend_normalizer.normalizers.extend(custom_normalizers_to_add)
        # Re-assign if `extend` doesn't modify in place or if Sequence needs explicit update
        # For `tokenizers.normalizers.Sequence`, `normalizers` is a list and `extend` modifies it in place.
        # No explicit re-assignment of tokenizer.backend_tokenizer.normalizer needed here unless Sequence is immutable.
        # To be safe, one might re-create:
        # tokenizer.backend_tokenizer.normalizer = Sequence(current_backend_normalizer.normalizers)
    else: # It's a single normalizer object, not a Sequence
        tokenizer.backend_tokenizer.normalizer = Sequence([current_backend_normalizer] + custom_normalizers_to_add)
# --- End Model & Tokenizer Configuration ---

title_md = """
<h1 style="text-align: center; margin-bottom: 5px;">AI Text Detector</h1>
<p style="text-align: center; font-size: 0.9em; color: var(--text-secondary); margin-top: 0; margin-bottom: 20px;">Developed by SzegedAI</p>
"""

description = """
<div class="app-description">
    <p>This tool utilizes the <b>ModernBERT</b> model to decide whether a given text is human-authored or AI-generated. It employs a soft voting ensemble of <b>three</b> models to improve detection accuracy.</p>
    <ul class="features-list">
        <li><span class="icon">✅</span> <strong>Human Verification: </strong> Clearly identifies human-written content.</li>
        <li><span class="icon">🔍</span> <strong>Model Detection: </strong> Capable of identifying content from over 40 AI models.</li>
        <li><span class="icon">📈</span> <strong>Accuracy: </strong> Performs optimally with more extensive text inputs.</li>
        <li><span class="icon">📄</span> <strong>Read more: </strong> Our methodology is detailed in our research paper: &nbsp;
            <a href="https://aclanthology.org/2025.genaidetect-1.15/" target="_blank" class="learn-more-link"> <b> LINK </b></a>.
        </li>
    </ul>
    <p class="instruction-text">Paste your text into the field below to analyze its origin.</p>
</div>
"""
bottom_text = "<p class='footer-text'>SzegedAI - Mihaly Kiss</p>"

AI_texts = [
    "Camels are remarkable desert animals known for their unique adaptations to harsh, arid environments. Native to the Middle East, North Africa, and parts of Asia, camels have been essential to human life for centuries, serving as a mode of transportation, a source of food, and even a symbol of endurance and survival. There are two primary species of camels: the dromedary camel, which has a single hump and is commonly found in the Middle East and North Africa, and the Bactrian camel, which has two humps and is native to Central Asia. Their humps store fat, not water, as commonly believed, allowing them to survive long periods without food by metabolizing the stored fat for energy. Camels are highly adapted to desert life. They can go for weeks without water, and when they do drink, they can consume up to 40 gallons in one sitting. Their thick eyelashes, sealable nostrils, and wide, padded feet protect them from sand and help them walk easily on loose desert terrain.",
]

Human_texts = [
    "To make BERT handle a variety of down-stream tasks, our input representation is able to unambiguously represent both a single sentence and a pair of sentences (e.g., h Question, Answeri) in one token sequence. Throughout this work, a “sentence” can be an arbitrary span of contiguous text, rather than an actual linguistic sentence. A “sequence” refers to the input token sequence to BERT, which may be a single sentence or two sentences packed together. We use WordPiece embeddings (Wu et al., 2016) with a 30,000 token vocabulary. The first token of every sequence is always a special classification token ([CLS]). The final hidden state corresponding to this token is used as the aggregate sequence representation for classification tasks. Sentence pairs are packed together into a single sequence."
]

def classify_text_interface(text):
    if not all([tokenizer, model_1, model_2, model_3]):
        return "<p style='text-align: center; color: var(--ai-color);'><strong>Error: Models not loaded. Please check the console.</strong></p>"

    cleaned_text = clean_text(text)
    if not cleaned_text.strip():
        result_message = "<p style='text-align: center; color: var(--text-secondary);'>Please enter some text to analyze.</p>"
        return result_message

    inputs = tokenizer(cleaned_text, return_tensors="pt", truncation=True, padding=True, max_length=512).to(device)

    with torch.no_grad():
        logits_1 = model_1(**inputs).logits
        logits_2 = model_2(**inputs).logits
        logits_3 = model_3(**inputs).logits

        softmax_1 = torch.softmax(logits_1, dim=1)
        softmax_2 = torch.softmax(logits_2, dim=1)
        softmax_3 = torch.softmax(logits_3, dim=1)

        averaged_probabilities = (softmax_1 + softmax_2 + softmax_3) / 3
        probabilities = averaged_probabilities[0]

    ai_probs = probabilities.clone()
    human_label_index = -1
    for k, v in label_mapping.items():
        if v.lower() == 'human':
            human_label_index = k
            break

    if human_label_index != -1:
        ai_probs[human_label_index] = 0
        human_prob_value = probabilities[human_label_index].item() * 100
    else:
        human_prob_value = 0
        print("Warning: 'human' label not found in label_mapping.")

    ai_total_prob = ai_probs.sum().item() * 100


    ai_argmax_index = torch.argmax(ai_probs).item()
    ai_argmax_model = label_mapping.get(ai_argmax_index, "Unknown AI")

    if human_prob_value > ai_total_prob :
        result_message = (
            f"<p><strong>The text is</strong> <span class='highlight-human'><strong>{human_prob_value:.2f}%</strong> likely <b>Human written</b>.</span></p>"
        )
    else:
        result_message = (
            f"<p><strong>The text is</strong> <span class='highlight-ai'><strong>{ai_total_prob:.2f}%</strong> likely <b>AI generated</b>.</span></p>"
            f"<p style='margin-top: 10px; font-size: 0.95em;'><strong>Most Likely AI Source:</strong> {ai_argmax_model} (with {probabilities[ai_argmax_index].item()*100:.2f}% confidence among AI models)</p>"
        )
    return result_message

modern_css = """
@import url('https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700&display=swap');

/* Define values for light and dark themes */
:root {
    --primary-bg-light: #F4F7FC;
    --app-bg-light: #FFFFFF;
    --text-primary-light: #2C3E50;
    --text-secondary-light: #7F8C8D;
    --accent-color-light: #1ABC9C;
    --accent-color-darker-light: #16A085;
    --border-color-light: #E0E6ED;
    --input-bg-light: #FFFFFF;
    --human-color-light: #2ECC71;
    --human-bg-light: rgba(46, 204, 113, 0.1);
    --ai-color-light: #E74C3C;
    --ai-bg-light: rgba(231, 76, 60, 0.1);
    --shadow-color-light: rgba(44, 62, 80, 0.1);
    --examples-bg-light: #F8F9FA;
    --placeholder-color-light: #B0BEC5;
    --accordion-label-color-light: var(--text-primary-light);
    --accordion-bg-light: var(--app-bg-light);
    --accordion-border-light: var(--border-color-light);
    --sample-textbox-bg-light: var(--input-bg-light);

    --primary-bg-dark: #121828; /* Even darker body for more contrast */
    --app-bg-dark: #1B2134; /* Darker app container */
    --text-primary-dark: #E0E7FF; /* Lighter text for dark mode */
    --text-secondary-dark: #98A0B3; /* Softer secondary text */
    --accent-color-dark: #2CE1C7; /* Brighter accent */
    --accent-color-darker-dark: #15B8A5;
    --border-color-dark: #2F364D; /* Subtle borders */
    --input-bg-dark: #22283E; /* Dark input fields */
    --human-color-dark: #50FA7B; /* Brighter lime green */
    --human-bg-dark: rgba(80, 250, 123, 0.15);
    --ai-color-dark: #FF79C6; /* Bright pink/magenta for AI */
    --ai-bg-dark: rgba(255, 121, 198, 0.15);
    --shadow-color-dark: rgba(0, 0, 0, 0.3); /* Shadow for dark mode */
    --examples-bg-dark: #22283E; /* Examples bg same as input */
    --placeholder-color-dark: #687083;
    --accordion-label-color-dark: var(--text-primary-dark);
    --accordion-bg-dark: var(--app-bg-dark);
    --accordion-border-dark: var(--border-color-dark);
    --sample-textbox-bg-dark: var(--input-bg-dark);

    /* Default to light theme variables */
    --primary-bg: var(--primary-bg-light);
    --app-bg: var(--app-bg-light);
    --text-primary: var(--text-primary-light);
    --text-secondary: var(--text-secondary-light);
    --accent-color: var(--accent-color-light);
    --accent-color-darker: var(--accent-color-darker-light);
    --border-color: var(--border-color-light);
    --input-bg: var(--input-bg-light);
    --input-focus-border: var(--accent-color-light); /* Default focus for light */
    --human-color: var(--human-color-light);
    --human-bg: var(--human-bg-light);
    --ai-color: var(--ai-color-light);
    --ai-bg: var(--ai-bg-light);
    --shadow-color: var(--shadow-color-light);
    --examples-bg: var(--examples-bg-light);
    --placeholder-color: var(--placeholder-color-light);
    --accordion-label-color: var(--accordion-label-color-light);
    --accordion-bg: var(--accordion-bg-light);
    --accordion-border: var(--accordion-border-light);
    --sample-textbox-bg: var(--sample-textbox-bg-light);

    --container-max-width: 800px;
    --border-radius-md: 8px;
    --border-radius-lg: 12px;
}

/* Apply Dark Theme when html.dark class is present (Hugging Face Spaces) */
html.dark {
    --primary-bg: var(--primary-bg-dark);
    --app-bg: var(--app-bg-dark);
    --text-primary: var(--text-primary-dark);
    --text-secondary: var(--text-secondary-dark);
    --accent-color: var(--accent-color-dark);
    --accent-color-darker: var(--accent-color-darker-dark);
    --border-color: var(--border-color-dark);
    --input-bg: var(--input-bg-dark);
    --input-focus-border: var(--accent-color-dark); /* Focus for dark */
    --human-color: var(--human-color-dark);
    --human-bg: var(--human-bg-dark);
    --ai-color: var(--ai-color-dark);
    --ai-bg: var(--ai-bg-dark);
    --shadow-color: var(--shadow-color-dark);
    --examples-bg: var(--examples-bg-dark);
    --placeholder-color: var(--placeholder-color-dark);
    --accordion-label-color: var(--accordion-label-color-dark);
    --accordion-bg: var(--accordion-bg-dark);
    --accordion-border: var(--accordion-border-dark);
    --sample-textbox-bg: var(--sample-textbox-bg-dark);
}

/* Fallback for system preference if html.dark is not set */
@media (prefers-color-scheme: dark) {
    html:not(.dark) :root { /* Apply only if HF class is not already active */
        --primary-bg: var(--primary-bg-dark);
        --app-bg: var(--app-bg-dark);
        --text-primary: var(--text-primary-dark);
        --text-secondary: var(--text-secondary-dark);
        --accent-color: var(--accent-color-dark);
        --accent-color-darker: var(--accent-color-darker-dark);
        --border-color: var(--border-color-dark);
        --input-bg: var(--input-bg-dark);
        --input-focus-border: var(--accent-color-dark);
        --human-color: var(--human-color-dark);
        --human-bg: var(--human-bg-dark);
        --ai-color: var(--ai-color-dark);
        --ai-bg: var(--ai-bg-dark);
        --shadow-color: var(--shadow-color-dark);
        --examples-bg: var(--examples-bg-dark);
        --placeholder-color: var(--placeholder-color-dark);
        --accordion-label-color: var(--accordion-label-color-dark);
        --accordion-bg: var(--accordion-bg-dark);
        --accordion-border: var(--accordion-border-dark);
        --sample-textbox-bg: var(--sample-textbox-bg-dark);
    }
}

.features-list strong::after {
  content: " ";
  display: inline-block;
  width: 0.2em;
}

body {
    font-family: 'Inter', sans-serif;
    background: var(--primary-bg);
    color: var(--text-primary);
    margin: 0;
    padding: 20px;
    display: flex;
    justify-content: center;
    align-items: flex-start;
    min-height: 100vh;
    box-sizing: border-box;
    overflow-y: auto;
    transition: background-color 0.2s ease-out, color 0.2s ease-out;
}

.gradio-container {
    background-color: var(--app-bg);
    border-radius: var(--border-radius-lg);
    padding: clamp(25px, 5vw, 40px);
    box-shadow: 0 8px 25px var(--shadow-color);
    max-width: var(--container-max-width);
    width: 100%;
    margin: 20px auto;
    border: 1px solid var(--border-color); /* Add subtle border consistent with theme */
    transition: background-color 0.2s ease-out, box-shadow 0.2s ease-out, border-color 0.2s ease-out;
}

/* Reset Gradio default styles that might interfere */
.form.svelte-633qhp, .block.svelte-11xb1hd, .gradio-html .block, .gradio-markdown > *:first-child {
    background: none !important;
    border: none !important;
    box-shadow: none !important;
    padding: 0 !important; /* Reset padding if it causes issues */
    margin: 0 !important; /* Reset margin for Markdown wrapper */
}
/* Ensure Markdown text color inherits correctly */
.gradio-markdown p, .gradio-markdown ul, .gradio-markdown li, .gradio-markdown h1, .gradio-markdown h2 {
    color: inherit !important;
}
.gradio-markdown a {
    color: var(--accent-color) !important;
}
.gradio-markdown a:hover {
    color: var(--accent-color-darker) !important;
}


.app-description p {
    color: var(--text-secondary);
    font-size: clamp(14px, 2.5vw, 16px);
    line-height: 1.7;
    margin-bottom: 15px !important; /* Override Gradio's specific p margin */
}

.app-description .instruction-text {
    font-weight: 500;
    color: var(--text-primary);
    margin-top: 20px !important;
    text-align: center;
}

.features-list {
    list-style: none;
    padding-left: 0;
    margin: 20px 0 !important;
}

.features-list li {
    display: flex;
    align-items: center;
    font-size: clamp(14px, 2.5vw, 16px);
    color: var(--text-secondary);
    margin-bottom: 12px !important;
    line-height: 1.6;
}

.features-list .icon {
    margin-right: 12px;
    font-size: 1.2em;
    color: var(--accent-color);
    flex-shrink: 0;
}

#text_input_box textarea {
    background-color: var(--input-bg);
    border: 1px solid var(--border-color);
    border-radius: var(--border-radius-md);
    font-size: clamp(15px, 2.5vw, 16px);
    padding: 15px;
    width: 100%;
    box-sizing: border-box;
    color: var(--text-primary);
    transition: background-color 0.2s ease-out, border-color 0.2s ease-out, box-shadow 0.2s ease-out, color 0.2s ease-out;
    min-height: 120px;
    box-shadow: 0 1px 3px rgba(0,0,0,0.03); /* Softer shadow */
}

#text_input_box textarea::placeholder {
    color: var(--placeholder-color);
    transition: color 0.2s ease-out;
}

#text_input_box textarea:focus {
    border-color: var(--input-focus-border);
    box-shadow: 0 0 0 3px color-mix(in srgb, var(--input-focus-border) 20%, transparent);
    outline: none;
}

#result_output_box {
    background-color: var(--input-bg);
    border: 1px solid var(--border-color);
    border-radius: var(--border-radius-md);
    padding: 20px;
    margin-top: 25px !important; /* Override Gradio */
    width: 100%;
    box-sizing: border-box;
    text-align: center;
    font-size: clamp(16px, 3vw, 17px);
    box-shadow: 0 1px 3px rgba(0,0,0,0.03);
    min-height: 80px;
    display: flex;
    flex-direction: column;
    justify-content: center;
    transition: background-color 0.2s ease-out, border-color 0.2s ease-out, color 0.2s ease-out;
}
#result_output_box p {
    margin-bottom: 8px !important;
    line-height: 1.6;
    color: var(--text-primary) !important;
}
#result_output_box p:last-child {
    margin-bottom: 0 !important;
}
#result_output_box strong {
    color: var(--text-primary) !important;
}

.highlight-human, .highlight-ai {
    font-weight: 600;
    padding: 5px 10px;
    border-radius: var(--border-radius-md);
    display: inline-block;
    font-size: 1.05em;
    transition: background-color 0.2s ease-out, color 0.2s ease-out;
}
.highlight-human { color: var(--human-color); background-color: var(--human-bg); }
.highlight-ai { color: var(--ai-color); background-color: var(--ai-bg); }

.gr-accordion {
    border: 1px solid var(--accordion-border) !important;
    border-radius: var(--border-radius-lg) !important;
    box-shadow: none !important;
    padding: 0 15px 15px 15px !important;
    margin-bottom: 20px !important;
    background-color: var(--accordion-bg) !important;
    transition: background-color 0.2s ease-out, border-color 0.2s ease-out;
}
.gr-accordion > .label-wrap button {
    font-weight: 600 !important;
    color: var(--accordion-label-color) !important;
    padding: 15px 0px !important;
    font-size: 1.05em !important;
    transition: color 0.2s ease-out;
}
.gr-accordion > .label-wrap { border-bottom: none !important; }

.gr-examples {
    padding: 15px 0px 0px 0px !important;
    border: none !important;
    border-radius: 0 !important;
    background-color: transparent !important;
    margin-top: 0px !important;
}
.gr-sample-textbox {
    border: 1px solid var(--border-color) !important;
    border-radius: var(--border-radius-md) !important;
    font-size: 14px !important;
    background-color: var(--sample-textbox-bg) !important;
    color: var(--text-primary) !important;
    transition: background-color 0.2s ease-out, border-color 0.2s ease-out, color 0.2s ease-out;
}
.gr-sample-textbox:hover { border-color: var(--accent-color) !important; }

.footer-text, #bottom_text {
    text-align: center;
    margin-top: 40px !important;
    font-size: clamp(13px, 2vw, 14px);
    color: var(--text-secondary);
}
#bottom_text p { margin: 0 !important; }

@media (max-width: 768px) {
    body { padding: 10px; align-items: flex-start; }
    .gradio-container { padding: 20px; margin: 10px; }
    /* h1 { font-size: 22px; }  Handled by Markdown inline style which uses clamp */
    .app-description p, .features-list li { font-size: 14px; }
    #text_input_box textarea { font-size: 15px; min-height: 100px; }
    #result_output_box { font-size: 15px; padding: 15px; }
    .gr-accordion > .label-wrap button { padding: 12px 0 !important; }
}
"""

iface = gr.Blocks(css=modern_css, theme=gr.themes.Base(font=[gr.themes.GoogleFont("Inter"), "sans-serif"]))

with iface:
    gr.Markdown(title_md)
    gr.Markdown(description)

    text_input = gr.Textbox(
        label="",
        placeholder="Type or paste your content here...",
        elem_id="text_input_box",
        lines=10
    )
    result_output = gr.HTML(elem_id="result_output_box")

    if all([tokenizer, model_1, model_2, model_3]):
        text_input.change(classify_text_interface, inputs=text_input, outputs=result_output)
    else:
        gr.HTML("<div id='result_output_box'><p style='color: var(--ai-color); text-align: center;'><strong>Application Error: Models could not be loaded. Please check the server console for details.</strong></p></div>")

    with gr.Accordion("AI Text Examples", open=False):
        gr.Examples(
            examples=AI_texts,
            inputs=text_input,
            label="",
        )

    with gr.Accordion("Human Text Examples", open=False):
        gr.Examples(
            examples=Human_texts,
            inputs=text_input,
            label="",
        )

    gr.Markdown(bottom_text, elem_id="bottom_text")

if __name__ == "__main__":
    iface.launch(share=False)