app.py

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from sentence_transformers import SentenceTransformer, util
import numpy as np
import requests
import gradio.themes as grthemes
import config

# ----------------------
# Paraphrasing Model Setup
# ----------------------
PARAPHRASE_MODEL_NAME = "Vamsi/T5_Paraphrase_Paws"
paraphrase_tokenizer = AutoTokenizer.from_pretrained(PARAPHRASE_MODEL_NAME)
paraphrase_model = AutoModelForSeq2SeqLM.from_pretrained(PARAPHRASE_MODEL_NAME)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
paraphrase_model = paraphrase_model.to(device)

# ----------------------
# Semantic Similarity Model
# ----------------------
similarity_model = SentenceTransformer('all-MiniLM-L6-v2')

# ----------------------
# Tone Templates
# ----------------------
tone_templates = {
    "Academic": "Paraphrase the following text in a formal, academic tone:",
    "Casual": "Paraphrase the following text in a casual, conversational tone:",
    "Friendly": "Paraphrase the following text in a friendly, approachable tone:",
    "Stealth": "Paraphrase the following text to bypass AI detectors and sound as human as possible:",
}

# ----------------------
# Paraphrasing Function
# ----------------------
def paraphrase(text, tone):
    prompt = tone_templates[tone] + " " + text
    input_ids = paraphrase_tokenizer.encode(prompt, return_tensors="pt", max_length=256, truncation=True).to(device)
    outputs = paraphrase_model.generate(
        input_ids,
        do_sample=True,
        top_k=120,
        top_p=0.95,
        temperature=0.7,
        repetition_penalty=1.2,
        max_length=256,
        num_return_sequences=1
    )
    paraphrased = paraphrase_tokenizer.decode(outputs[0], skip_special_tokens=True)
    return paraphrased

# ----------------------
# Semantic Similarity Function
# ----------------------
def semantic_similarity(text1, text2):
    emb1 = similarity_model.encode(text1, convert_to_tensor=True)
    emb2 = similarity_model.encode(text2, convert_to_tensor=True)
    sim = util.pytorch_cos_sim(emb1, emb2).item()
    return sim

# ----------------------
# Real AI Detection (Winston AI API)
# ----------------------
def check_ai_score(text):
    api_key = config.WINSTON_AI_API_KEY
    api_url = config.WINSTON_AI_API_URL
    if not api_key:
        return None, "No API key set. Please add your Winston AI API key to config.py."
    headers = {
        "Authorization": f"Bearer {api_key}",
        "Content-Type": "application/json"
    }
    data = {"text": text, "sentences": False}
    try:
        response = requests.post(api_url, headers=headers, json=data, timeout=30)
        if response.status_code == 200:
            result = response.json()
            # Winston AI returns a 'score' (0-100, higher = more human)
            score = result.get("score", None)
            if score is not None:
                ai_prob = 1.0 - (score / 100.0)
                return ai_prob, None
            else:
                return None, "No score in Winston AI response."
        else:
            return None, f"Winston AI error: {response.status_code} {response.text}"
    except Exception as e:
        return None, f"Winston AI exception: {str(e)}"

# ----------------------
# Humanization Score & Rating
# ----------------------
def humanization_score(sim, ai_prob):
    # Lower similarity and lower AI probability = more human
    score = (1.0 - sim) * 0.5 + (1.0 - ai_prob) * 0.5
    return score

def humanization_rating(score):
    if score < 0.7:
        return f"⚠️ Still AI-like ({score:.2f})"
    elif score < 0.85:
        return f"👍 Acceptable ({score:.2f})"
    else:
        return f"✅ Highly Humanized ({score:.2f})"

# ----------------------
# Main Processing Function
# ----------------------
def process(text, tone):
    if not text.strip():
        return "", "", 0.0, "", 0.0, ""
    # Pre-humanization AI detection
    pre_ai_prob, pre_err = check_ai_score(text)
    if pre_ai_prob is None:
        return "", f"AI Detection Error: {pre_err}", 0.0, "", 0.0, ""
    # Paraphrase
    try:
        paraphrased = paraphrase(text, tone)
    except Exception as e:
        return f"[Paraphrasing error: {str(e)}]", "", 0.0, "", 0.0, ""
    # Post-humanization AI detection
    post_ai_prob, post_err = check_ai_score(paraphrased)
    if post_ai_prob is None:
        return paraphrased, f"AI Detection Error: {post_err}", 0.0, "", 0.0, ""
    # Semantic similarity
    sim = semantic_similarity(text, paraphrased)
    # Humanization score
    score = humanization_score(sim, post_ai_prob)
    rating = humanization_rating(score)
    ai_score_str = f"Pre: {100*(1-pre_ai_prob):.1f}% human | Post: {100*(1-post_ai_prob):.1f}% human"
    return (
        paraphrased,         # gr.Textbox (string)
        ai_score_str,        # gr.Markdown (string)
        sim,                # gr.Number (float)
        rating,             # gr.Markdown (string)
        score * 100,        # gr.Number (float)
        ""
    )

# ----------------------
# Gradio UI
# ----------------------
custom_theme = grthemes.Base(
    primary_hue="blue",
    secondary_hue="blue",
    neutral_hue="slate"
)

with gr.Blocks(theme=custom_theme, title="AI Humanizer - Made by Taha") as demo:
    gr.Markdown("""
    # 🧠 AI Humanizer
    <div style='display:flex;justify-content:space-between;align-items:center;'>
        <span style='font-size:1.2em;color:#7bb1ff;'>Rewrite AI text to sound 100% human</span>
        <span style='font-weight:bold;color:#7bb1ff;'>Made by Taha</span>
    </div>
    """, elem_id="header")
    with gr.Row():
        with gr.Column():
            text_in = gr.Textbox(label="Paste AI-generated text here", lines=8, placeholder="Paste your text...", elem_id="input-box")
            tone = gr.Dropdown(["Academic", "Casual", "Friendly", "Stealth"], value="Stealth", label="Tone Selector")
            btn = gr.Button("Humanize", elem_id="humanize-btn")
        with gr.Column():
            text_out = gr.Textbox(label="Humanized Output", lines=8, interactive=False, elem_id="output-box")
            ai_scores = gr.Markdown("", elem_id="ai-scores")
            sim_score = gr.Number(label="Similarity (0=very different, 1=very similar)", interactive=False)
            rating = gr.Markdown("", elem_id="rating")
            human_score = gr.Number(label="Humanization Score (%)", interactive=False)
    btn.click(
        process,
        inputs=[text_in, tone],
        outputs=[text_out, ai_scores, sim_score, rating, human_score, gr.Textbox(visible=False)],
        api_name="humanize"
    )
    gr.Markdown("""
    <div style='text-align:center;color:#7bb1ff;margin-top:2em;'>
        <b>Made by Taha</b> | Free for unlimited use | Optimized for students and creators
    </div>
    """, elem_id="footer")

demo.launch()