app.py

from spacy.lang.en import English
from spacy.lang.en.stop_words import STOP_WORDS
import pandas as pd
import gradio as gr
from transformers import pipeline
from gradio.themes.utils.colors import red, green
import requests
import json
import os
from dotenv import load_dotenv
import time

# Load environment variables
load_dotenv()

# Initialize the NLP pipeline
nlp = English()
nlp.add_pipe("sentencizer")
tokenizer = nlp.tokenizer

# Initialize the text classification pipeline
detector = pipeline(task='text-classification', model='SJTU-CL/RoBERTa-large-ArguGPT-sent')

# Groq API configuration
GROQ_API_KEY = os.getenv("GROQ_API_KEY")
if not GROQ_API_KEY:
    raise ValueError("Please set your GROQ_API_KEY in the .env file")
GROQ_API_URL = "https://api.groq.com/openai/v1/chat/completions"
GROQ_MODEL = "llama3-70b-8192"  # Updated to latest model

# Define color map for highlighted text
color_map = {
    '0%': green.c400,
    '10%': green.c300,
    '20%': green.c200,
    '30%': green.c100,
    '40%': green.c50,
    '50%': red.c50,
    '60%': red.c100,
    '70%': red.c200,
    '80%': red.c300,
    '90%': red.c400,
    '100%': red.c500
}

def is_stopword(word):
    """Check if a word is a stop word or very short"""
    return word.lower() in STOP_WORDS or len(word) <= 2

def get_synonyms(word):
    """Get simple, human-readable synonyms using Groq API"""
    if is_stopword(word):
        return [word]  # Don't get synonyms for stop words
    
    headers = {
        "Authorization": f"Bearer {GROQ_API_KEY}",
        "Content-Type": "application/json"
    }
    
    prompt = f"""Provide a list of exactly 5 simple synonyms for '{word}'. 
    Return ONLY a JSON array of synonyms without any additional text. 
    Example: ["use", "employ", "apply", "make use of", "take"]"""
    
    data = {
        "model": GROQ_MODEL,
        "messages": [{"role": "user", "content": prompt}],
        "temperature": 0.3,
        "max_tokens": 100,
        "response_format": {"type": "json_object"}
    }
    
    try:
        response = requests.post(GROQ_API_URL, headers=headers, json=data)
        response.raise_for_status()
        result = response.json()
        content = json.loads(result['choices'][0]['message']['content'])
        
        # Improved response parsing
        if isinstance(content, dict):
            for key in ['synonyms', 'words', 'alternatives']:
                if key in content and isinstance(content[key], list):
                    return content[key][:5]
        return [word]  # Fallback if parsing fails
    except Exception as e:
        print(f"Error getting synonyms: {e}")
        return [word]  # Fallback to original word if API fails

def identify_problem_words(text):
    """Use Groq API to identify uncommon, difficult, and AI-generated words"""
    headers = {
        "Authorization": f"Bearer {GROQ_API_KEY}",
        "Content-Type": "application/json"
    }
    
    prompt = f"""Analyze this text and return ONLY a JSON list of words that are:
    1. Uncommon (not in everyday vocabulary)
    2. Difficult (complex or technical)
    3. Likely AI-generated (overly formal, verbose, or unnatural)
    
    Exclude all stop words (a, an, the, and, but, etc.) and very short words (1-2 letters).
    Return format: {{"words": ["word1", "word2", ...]}}
    
    Text: {text}"""
    
    data = {
        "model": GROQ_MODEL,
        "messages": [{"role": "user", "content": prompt}],
        "temperature": 0.2,
        "max_tokens": 200,
        "response_format": {"type": "json_object"}
    }
    
    try:
        response = requests.post(GROQ_API_URL, headers=headers, json=data)
        response.raise_for_status()
        result = response.json()
        content = json.loads(result['choices'][0]['message']['content'])
        
        if isinstance(content, dict) and 'words' in content:
            # Filter out any stop words that might have slipped through
            filtered_words = [word for word in content['words'] if not is_stopword(word)]
            return set(filtered_words)
        return set()
    except Exception as e:
        print(f"Error identifying problem words: {e}")
        return set()

def predict_word(word, problem_words):
    """Predict AI probability for a single word if it's in problem words"""
    if len(word) <= 3 or word.lower() not in problem_words or is_stopword(word):
        return 0.0
    try:
        prob = predict_one_sent(word)
        return prob
    except:
        return 0.0

def predict_doc(doc):
    start_time = time.time()
    
    # First identify problem words using Groq
    problem_words = identify_problem_words(doc)
    print(f"Identified problem words: {problem_words}")
    
    sents = [s.text for s in nlp(doc).sents]
    data = {'sentence': [], 'label': [], 'score': []}
    sent_res = []
    word_highlights = []
    
    for sent in sents:
        sent_prob = predict_one_sent(sent)
        
        # Word-level analysis - only for problem words
        tokens = [token.text for token in tokenizer(sent)]
        word_probs = [predict_word(token, problem_words) for token in tokens]
        
        for word, prob in zip(tokens, word_probs):
            if prob >= 0.2:  # Only highlight words with >20% AI probability
                if prob < 0.3: label = '20%'
                elif prob < 0.4: label = '30%'
                elif prob < 0.5: label = '40%'
                elif prob < 0.6: label = '50%'
                elif prob < 0.7: label = '60%'
                elif prob < 0.8: label = '70%'
                elif prob < 0.9: label = '80%'
                elif prob < 1: label = '90%'
                else: label = '100%'
                word_highlights.append((word, label))
            else:
                word_highlights.append((word, None))

        data['sentence'].append(sent)
        data['score'].append(round(sent_prob, 4))
        if sent_prob <= 0.5:
            data['label'].append('Human')
        else:
            data['label'].append('Machine')

        if sent_prob < 0.1: label = '0%'
        elif sent_prob < 0.2: label = '10%'
        elif sent_prob < 0.3: label = '20%'
        elif sent_prob < 0.4: label = '30%'
        elif sent_prob < 0.5: label = '40%'
        elif sent_prob < 0.6: label = '50%'
        elif sent_prob < 0.7: label = '60%'
        elif sent_prob < 0.8: label = '70%'
        elif sent_prob < 0.9: label = '80%'
        elif sent_prob < 1: label = '90%'
        else: label = '100%'
        sent_res.append((sent, label))

    df = pd.DataFrame(data)
    csv_path = 'result.csv'
    df.to_csv(csv_path)
    print(f"Analysis took {time.time() - start_time:.2f} seconds")
    
    overall_score = df.score.mean()
    overall_label = 'Human' if overall_score <= 0.5 else 'Machine'
    sum_str = f'The essay is probably written by {overall_label}. The probability of being generated by AI is {overall_score:.2f}'

    return sum_str, sent_res, df, csv_path, word_highlights

def predict_one_sent(sent):
    res = detector(sent)[0]
    org_label, prob = res['label'], res['score']
    if org_label == 'LABEL_0': prob = 1 - prob
    return prob

def update_text(text, selected_word, replacement, word_highlights):
    new_text = text.replace(selected_word, replacement, 1)
    
    # Update word_highlights with the new word (assuming it's now human-written)
    updated_highlights = []
    replaced = False
    for word, label in word_highlights:
        if word == selected_word and not replaced:
            updated_highlights.append((replacement, '0%'))
            replaced = True
        else:
            updated_highlights.append((word, label))
    
    return new_text, updated_highlights

def process_word_highlights(highlights):
    return highlights

# Custom CSS for modern look
custom_css = """
.gradio-container {
    font-family: 'Arial', sans-serif;
}
.gradio-header {
    background-color: #4CAF50;
    color: white;
    padding: 10px;
    text-align: center;
    border-radius: 8px;
    margin-bottom: 20px;
}
.gradio-button {
    background-color: #4CAF50;
    color: white;
    border: none;
    padding: 10px 20px;
    text-align: center;
    text-decoration: none;
    display: inline-block;
    font-size: 16px;
    margin: 4px 2px;
    cursor: pointer;
    border-radius: 5px;
    transition: background-color 0.3s;
}
.gradio-button:hover {
    background-color: #45a049;
}
.highlighted-word {
    cursor: pointer;
    padding: 2px 4px;
    border-radius: 3px;
    transition: all 0.2s;
}
.highlighted-word:hover {
    text-decoration: underline;
    background-color: #f0f0f0;
    transform: scale(1.05);
}
.replacement-row {
    border: 1px solid #ddd;
    padding: 15px;
    border-radius: 8px;
    margin-top: 10px;
    background-color: #f9f9f9;
}
"""

with gr.Blocks(css=custom_css) as demo:
    gr.Markdown("""## AI vs Human Essay Detector""")
    gr.Markdown("""Identify and replace uncommon, difficult, and AI-generated words in your text.""")
    
    word_highlights = gr.State([])
    selected_word = gr.State("")
    
    with gr.Row():
        with gr.Column():
            text_in = gr.Textbox(
                lines=10, 
                label='Essay Input', 
                placeholder="Paste your essay here...",
                elem_classes=["text-input"]
            )
            btn = gr.Button('Analyze Text', variant="primary")

        with gr.Column():
            sent_res = gr.HighlightedText(
                label='Sentence-level Analysis', 
                color_map=color_map,
                show_legend=True
            )
            word_res = gr.HighlightedText(
                label='Word-level Analysis (Click words to replace)', 
                color_map=color_map,
                show_legend=True
            )
    
    with gr.Row():
        summary = gr.Textbox(label='Overall Analysis', interactive=False)
        csv_f = gr.File(label='Download Detailed Analysis')

    with gr.Row():
        tab = gr.Dataframe(
            label='Detailed Sentence Analysis', 
            wrap=True,
            max_rows=10
        )
        
    with gr.Column(visible=False) as replacement_row:
        gr.Markdown("### Replace Word")
        with gr.Row():
            replacement_dropdown = gr.Dropdown(
                label="Select replacement",
                interactive=True,
                allow_custom_value=True
            )
        with gr.Row():
            replace_btn = gr.Button("Replace", variant="primary")
            cancel_btn = gr.Button("Cancel")
    
    def on_word_select(evt: gr.SelectData):
        if evt.value:
            synonyms = get_synonyms(evt.value)
            return (
                evt.value, 
                gr.Dropdown(choices=synonyms, value=evt.value), 
                gr.Column(visible=True)
            )
        return None, None, gr.Column(visible=False)
    
    word_res.select(
        fn=on_word_select,
        outputs=[selected_word, replacement_dropdown, replacement_row]
    )
    
    replace_btn.click(
        fn=update_text,
        inputs=[text_in, selected_word, replacement_dropdown, word_highlights],
        outputs=[text_in, word_highlights]
    ).then(
        fn=lambda: gr.Column(visible=False),
        outputs=replacement_row
    ).then(
        fn=lambda x: predict_doc(x),
        inputs=text_in,
        outputs=[summary, sent_res, tab, csv_f, word_highlights]
    ).then(
        fn=process_word_highlights,
        inputs=word_highlights,
        outputs=word_res
    )
    
    cancel_btn.click(
        fn=lambda: gr.Column(visible=False),
        outputs=replacement_row
    )
    
    btn.click(
        fn=predict_doc,
        inputs=text_in,
        outputs=[summary, sent_res, tab, csv_f, word_highlights]
    ).then(
        fn=process_word_highlights,
        inputs=word_highlights,
        outputs=word_res
    )

if __name__ == "__main__":
    demo.launch()