app.py

# app.py — Text detector with image-app style output (40-word min, loader, HTML result)
import os, sys, traceback
import gradio as gr
import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from huggingface_hub import HfApi
from huggingface_hub.utils import RepositoryNotFoundError, HfHubHTTPError

# -------- CONFIG --------
MODEL_ID = os.environ.get("MODEL_ID", "AICodexLab/answerdotai-ModernBERT-base-ai-detector")
HF_TOKEN = os.environ.get("HF_TOKEN", None)  # add as secret if model private
MIN_WORDS = int(os.environ.get("MIN_WORDS", 40))
device = "cuda" if torch.cuda.is_available() else "cpu"
auth_kwargs = {"use_auth_token": HF_TOKEN} if HF_TOKEN else {}

model = None
tokenizer = None
load_error = None

# -------- MODEL LOADING (safe) --------
def try_load_model():
    global model, tokenizer, load_error
    try:
        api = HfApi(token=HF_TOKEN) if HF_TOKEN else HfApi()
        _ = api.model_info(MODEL_ID)  # verify existence/access
        tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, **auth_kwargs)
        model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID, **auth_kwargs)
        model.to(device)
        model.eval()
        load_error = None
        print(f"[INFO] Model loaded: {MODEL_ID}", flush=True)
    except RepositoryNotFoundError as e:
        load_error = f"Repository not found or access denied: {MODEL_ID}. {e}"
        traceback.print_exc()
    except HfHubHTTPError as e:
        load_error = f"Hugging Face Hub HTTP error: {e}"
        traceback.print_exc()
    except Exception as e:
        load_error = f"Failed to load model {MODEL_ID}: {repr(e)}"
        traceback.print_exc()

try_load_model()

# -------- HELPERS --------
def count_words(text: str) -> int:
    if not text:
        return 0
    return len([w for w in text.strip().split() if w])

def make_result_html(verdict_text: str, color: str):
    """Return styled HTML block like your example."""
    html = f"""
    <div class='result-box' style="
        background: linear-gradient(135deg, {color}33, #1a1a1a);
        border: 2px solid {color};
        border-radius: 15px;
        padding: 22px;
        text-align: center;
        color: white;
        font-size: 20px;
        font-weight: 700;
        box-shadow: 0 0 20px {color}55;
        animation: fadeIn 0.6s ease-in-out;
    ">
        {verdict_text}
    </div>
    """
    return html

# -------- PREDICTION (generator to show loader then result) --------
def analyze_text(text: str):
    # loader HTML (pulse)
    loader_html = "<div id='pulse-loader'></div>"

    # show loader immediately
    yield (loader_html, "")

    # check model load
    if load_error:
        err_html = f"<div style='color:#ff4d4f;font-weight:700;'>Model load error: {load_error}</div>"
        yield ("", err_html)
        return

    if not text or text.strip() == "":
        yield ("", "<div style='color:red;'>Please enter some text first.</div>")
        return

    wc = count_words(text)
    if wc < MIN_WORDS:
        yield ("", f"<div style='color:#ff9900;font-weight:600;'>⚠️ Please enter at least {MIN_WORDS} words (currently {wc}).</div>")
        return

    try:
        inputs = tokenizer(text, truncation=True, padding=True, return_tensors="pt", max_length=512)
        inputs = {k: v.to(device) for k, v in inputs.items()}
        with torch.no_grad():
            out = model(**inputs)
            logits = out.logits
            probs = F.softmax(logits, dim=-1).cpu().numpy()[0]

        # interpret probs
        if probs.shape[0] == 1:
            ai_prob = float(probs[0])
            human_prob = 1.0 - ai_prob
        else:
            human_prob = float(probs[0])
            ai_prob = float(probs[1])

        # percent confidence (we'll show the predicted label's confidence)
        if ai_prob >= human_prob:
            label = "AI-generated"
            conf = ai_prob * 100.0
            color = "#007BFF"  # blue
        else:
            label = "Human"
            conf = human_prob * 100.0
            color = "#4CAF50"  # green

        verdict = f"{label} ({conf:.1f}% confidence)"
        html = make_result_html(verdict, color)
        yield ("", html)
    except Exception as e:
        traceback.print_exc()
        yield ("", f"<div style='color:red;'>Error analyzing text: {str(e)}</div>")

# -------- CSS (pulse loader + theme) --------
css = """
body, .gradio-container {
    font-family: 'Poppins', sans-serif !important;
    background: transparent !important;
}
h1 {
    text-align: center;
    font-weight: 700;
    color: #007BFF;
    margin-bottom: 10px;
}
.gr-button-primary {
    background-color: #007BFF !important;
    color: white !important;
    font-weight: 600;
    border-radius: 10px;
    height: 48px;
}
.gr-button-secondary {
    background-color: #dc3545 !important;
    color: white !important;
    border-radius: 10px;
    height: 48px;
}
#pulse-loader {
    width: 100%;
    height: 6px;
    background: linear-gradient(90deg, #007BFF, #00C3FF);
    animation: pulse 1.2s infinite ease-in-out;
    border-radius: 4px;
    box-shadow: 0 0 10px #007BFF;
    margin-top: 6px;
    margin-bottom: 6px;
}
@keyframes pulse {
    0% { transform: scaleX(0.05); opacity: 0.6; }
    50% { transform: scaleX(1); opacity: 1; }
    100% { transform: scaleX(0.05); opacity: 0.6; }
}
@keyframes fadeIn {
    from { opacity: 0; transform: scale(0.95); }
    to { opacity: 1; transform: scale(1); }
}
.result-box { /* fallback style for non-inline versions */ }
"""


# -------- GRADIO APP --------
with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
    gr.Markdown("<h1>🔎 AI Text Detector</h1>")
    gr.Markdown("Detect whether a given text is AI-generated or human-written using the ModernBERT model.")

    with gr.Row():
        with gr.Column(scale=1):
            txt = gr.Textbox(label="Enter your text (at least 40 words)", lines=10, placeholder="Paste your paragraph here...")
            analyze_btn = gr.Button("Analyze", variant="primary")
            clear_btn = gr.Button("Clear", variant="secondary")
            loader = gr.HTML("")  # loader animation
        with gr.Column(scale=1):
            result_html = gr.HTML(label="Result")

    # --- Analyze button ---
    analyze_btn.click(analyze_text, inputs=txt, outputs=[loader, result_html])

    # --- Clear button (fix: clears input + output + loader) ---
    def clear_all():
        return "", "", ""
    clear_btn.click(clear_all, outputs=[txt, loader, result_html])

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=int(os.environ.get("PORT", 7860)))