app.py

import os
import re
import unicodedata
import difflib
import gradio as gr
from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser

# ==========================================
# 1. CONFIGURATION & SECURITY
# ==========================================
OPENROUTER_KEY = os.environ.get('OPENROUTER_API_KEY')
BASE_URL = "https://openrouter.ai/api/v1"

APP_USER = os.environ.get('APP_USERNAME')
APP_PASS = os.environ.get('APP_PASSWORD')

# Speed-Optimized Model Selection with Fallbacks
MODEL_OPTIONS = {
    "⚡ Llama 3.1 8B (Ultra-Fast)": "meta-llama/llama-3.1-8b-instruct:free,openrouter/free",
    "⚡ Gemini 2.0 Flash Lite (Google's Fastest)": "google/gemini-2.0-flash-lite-001:free,openrouter/free",
    "⚡ Mistral Small 24B (Fast & Nuanced)": "mistralai/mistral-small-24b-instruct-2501:free,openrouter/free",
    "Step 3.5 Flash (Balanced)": "stepfun/step-3.5-flash:free,openrouter/free"
}

# ==========================================
# 2. THE PROMPTS (3-Pass System with Zero Data Loss Lock)
# ==========================================

# Pass 1: Structural Demolition
pass1_prompt = ChatPromptTemplate.from_template(
    "You are an expert editor. Rewrite the following text to break any predictable, formulaic AI structures and grammar footprints. "
    "CRITICAL DATA RULE: You must retain 100% of the original information. Do NOT drop a single fact, number, name, or nuanced detail. Do NOT summarize. "
    "1. SYNTACTIC SABOTAGE: Maximize sentence length variance (burstiness). Mix very short fragments with longer, complex sentences. "
    "2. POS SHIFT: AI models overuse active voice and linking words. You must occasionally use passive voice to break predictability. "
    "3. NO CLICHES: Delete all AI transition clichés like 'This isn't just X, it's Y', 'Furthermore', or 'It is worth noting'. "
    "Keep all original facts and data completely accurate, but scramble the logical flow so it reads less like a textbook. "
    "OUTPUT ONLY THE REWRITTEN TEXT. No introductions or explanations.\n\nText: {input}"
)

# Pass 2: Style & Guardrails
pass2_prompt = ChatPromptTemplate.from_template(
    "Edit this text to sound perfectly human, plain-spoken, and direct. "
    "Imagine you are writing a practical update to a smart colleague. "
    "CRITICAL RULES: "
    "1. ZERO DATA LOSS: Do not summarize or remove any information. Every fact, number, and entity from the draft must be perfectly preserved. "
    "2. STRICT OUTPUT: OUTPUT ABSOLUTELY NOTHING EXCEPT THE FINAL TEXT. Do not add notes or list changes. "
    "3. NO EMPTY FILLER: Never use meta-commentary like 'I have more to say, though'. Just state the facts. "
    "4. NO DRAMATICS: Absolutely no rhetorical questions. Avoid poetic metaphors and the 'Rule of Three'. "
    "5. NATURAL DATA: Never use the words 'respectively' or 'accounted for'. Blend numbers into the sentence naturally. "
    "6. PERPLEXITY INJECTION: Lower the vocabulary complexity slightly. Choose slightly less common synonyms to avoid predictable word pairings. "
    "7. PUNCTUATION: Absolutely NO m-dashes. Use standard hyphens, commas, periods, or parentheses instead. "
    "8. VOCABULARY: Do not use corporate fluff. "
    "Keep the tone grounded, highly efficient, and slightly informal.\n\nText: {input}"
)

# Pass 3: Flow & Cohesion
pass3_prompt = ChatPromptTemplate.from_template(
    "You are the final proofreader. Your job is to review this text for natural flow and cohesion without losing any data. "
    "CRITICAL RULES: "
    "1. ZERO DATA LOSS: You are strictly forbidden from deleting any facts, names, or numbers. Do not shorten or summarize the text. "
    "2. OUTPUT ONLY THE FINAL TEXT. No intros, no notes. "
    "3. Smooth out any choppy or awkward transitions left by previous edits. Ensure it reads naturally. "
    "4. Maintain all previous rules: No m-dashes, no corporate fluff, no filler phrases. "
    "Make it flow perfectly while preserving 100% of the information.\n\nText: {input}"
)

# ==========================================
# 3. PIPELINE FUNCTIONS
# ==========================================

def sanitize_text(text):
    """Text Normalization & Post-Processing Cleanup"""
    if not text:
        return ""
    
    cleaned = re.sub(r'[\u200B-\u200D\uFEFF\u200E\u200F\u202A-\u202E]', '', text)
    cleaned = unicodedata.normalize('NFKC', cleaned)
    
    homoglyphs = {
        'Α': 'A', 'Β': 'B', 'Ε': 'E', 'Η': 'H', 'Ι': 'I', 'Κ': 'K', 'Μ': 'M', 
        'Ν': 'N', 'Ο': 'O', 'Ρ': 'P', 'Τ': 'T', 'Χ': 'X', 'Υ': 'Y', 'Ζ': 'Z',
        'а': 'a', 'с': 'c', 'е': 'e', 'о': 'o', 'р': 'p', 'х': 'x', 'у': 'y',
        'А': 'A', 'В': 'B', 'С': 'C', 'Е': 'E', 'Н': 'H', 'О': 'O', 'Р': 'P', 
        'Т': 'T', 'Х': 'X', 'У': 'Y', 'і': 'i', 'ј': 'j'
    }
    trans_table = str.maketrans(homoglyphs)
    cleaned = cleaned.translate(trans_table)
    
    cleaned = cleaned.replace("—", "-").replace("–", "-")
    
    return cleaned.strip()

def calculate_similarity(original, final):
    """Semantic Similarity Check (Lexical Preservation)"""
    seq = difflib.SequenceMatcher(None, original.split(), final.split())
    ratio = seq.ratio() * 100
    
    if ratio > 80:
        status = "⚠️ Danger: Too similar to original AI text."
    elif ratio < 25:
        status = "⚠️ Warning: High variation. Verify that no data was dropped."
    else:
        status = "✅ Optimal variation achieved."
        
    return f"**Preservation Score:** {ratio:.1f}%\n\n{status}"

# ==========================================
# 4. MAIN EXECUTION
# ==========================================

def execute_pipeline(input_text, model_choice):
    # STEP 1: Text Normalization
    normalized_input = sanitize_text(input_text)
    
    if not normalized_input:
        return "Please enter some text to humanize.", ""
        
    try:
        selected_model_id = MODEL_OPTIONS[model_choice]
        
        active_model = ChatOpenAI(
            model=selected_model_id, 
            openai_api_key=OPENROUTER_KEY,
            openai_api_base=BASE_URL,
            temperature=0.85, 
            max_tokens=6000 
        )

        # Build the 3-Pass Chain
        pipeline_chain = (
            {"input": lambda x: x}
            | pass1_prompt       # STEP 2: Rewrite Pass 1 (Structure)
            | active_model 
            | StrOutputParser() 
            | (lambda text: {"input": text})
            | pass2_prompt       # STEP 3: Rewrite Pass 2 (Style)
            | active_model 
            | StrOutputParser()
            | (lambda text: {"input": text})
            | pass3_prompt       # STEP 4: Rewrite Pass 3 (Flow)
            | active_model
            | StrOutputParser()
        )

        # Run the LLM chain
        raw_result = pipeline_chain.invoke(normalized_input)
        
        # STEP 5: Post-processing Cleanup
        final_output = sanitize_text(raw_result)
        
        # STEP 6: Semantic Similarity Check
        similarity_metrics = calculate_similarity(normalized_input, final_output)
        
        # STEP 7: Final Output
        return final_output, similarity_metrics
        
    except Exception as e:
        return f"An error occurred. Check your API key or try again later.\nError details: {e}", ""

# UI Helper Functions
def update_word_count(text):
    count = len(text.split()) if text else 0
    return f"**Word Count:** {count} words"

def clear_boxes():
    return "", "", "**Word Count:** 0 words", "**Word Count:** 0 words", ""

# ==========================================
# 5. USER INTERFACE
# ==========================================

with gr.Blocks() as app:
    gr.Markdown("# 🤖 -> 🧑 The AI Humanizer (V2)")
    gr.Markdown("Passes your text through a 3-stage LLM pipeline (Structure > Style > Flow) while sanitizing watermarks.")
    
    with gr.Row():
        # LEFT COLUMN
        with gr.Column():
            model_dropdown = gr.Dropdown(
                choices=list(MODEL_OPTIONS.keys()), 
                value="⚡ Llama 3.1 8B (Ultra-Fast)", 
                label="Select AI Model"
            )
            
            input_box = gr.Textbox(
                lines=12, 
                placeholder="Paste your AI text here...", 
                label="Original AI Text"
            )
            input_word_count = gr.Markdown("**Word Count:** 0 words")
            
            with gr.Row():
                clear_btn = gr.Button("🗑️ Clear Text", variant="secondary")
                submit_btn = gr.Button("✨ Run Pipeline", variant="primary")
        
        # RIGHT COLUMN
        with gr.Column():
            output_box = gr.Textbox(
                lines=12, 
                label="Humanized Output"
            )
            
            with gr.Row():
                output_word_count = gr.Markdown("**Word Count:** 0 words")
                similarity_display = gr.Markdown("**Preservation Score:** N/A")
                
            copy_btn = gr.Button("📋 Copy Output", variant="secondary")
            
    # Connect the UI elements
    input_box.change(fn=update_word_count, inputs=input_box, outputs=input_word_count)
    output_box.change(fn=update_word_count, inputs=output_box, outputs=output_word_count)
    
    submit_btn.click(
        fn=execute_pipeline, 
        inputs=[input_box, model_dropdown], 
        outputs=[output_box, similarity_display]
    )
    
    clear_btn.click(
        fn=clear_boxes,
        inputs=[],
        outputs=[input_box, output_box, input_word_count, output_word_count, similarity_display]
    )
    
    copy_btn.click(
        fn=None,
        inputs=[output_box],
        outputs=None,
        js="(text) => { navigator.clipboard.writeText(text); }"
    )

# ==========================================
# 6. SECURE LAUNCH
# ==========================================
if APP_USER and APP_PASS:
    app.launch(auth=(APP_USER, APP_PASS), theme=gr.themes.Soft())
else:
    app.launch(theme=gr.themes.Soft())