app.py

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
import nltk
from datetime import datetime, timedelta
import requests
from bs4 import BeautifulSoup
import json

# Download required NLTK data
try:
    nltk.data.find('tokenizers/punkt')
except LookupError:
    nltk.download('punkt')

# Initialize models and tokenizers
def load_models():
    # Text generation model
    generator_model = "facebook/opt-350m"
    generator_tokenizer = AutoTokenizer.from_pretrained(generator_model)
    generator = AutoModelForCausalLM.from_pretrained(generator_model)
    
    # Sentiment analysis
    sentiment_analyzer = pipeline(
        "sentiment-analysis",
        model="finiteautomata/bertweet-base-sentiment-analysis"
    )
    
    # Bias detection
    bias_detector = pipeline(
        "text-classification",
        model="unitary/unbiased-bertscore"
    )
    
    return generator_tokenizer, generator, sentiment_analyzer, bias_detector

# Simplified news fetching function
def fetch_recent_news(query, num_articles=3):
    # Using Google News RSS feed
    base_url = "https://news.google.com/rss/search"
    params = {
        'q': query,
        'hl': 'en-US',
        'gl': 'US',
        'ceid': 'US:en'
    }
    
    try:
        response = requests.get(base_url, params=params, timeout=5)
        soup = BeautifulSoup(response.content, 'xml')
        items = soup.find_all('item', limit=num_articles)
        
        news_data = []
        for item in items:
            try:
                news_data.append({
                    'title': item.title.text,
                    'description': item.description.text if item.description else ""
                })
            except:
                continue
        
        return news_data
    except Exception as e:
        return [{'title': f'Using default context due to error: {str(e)}', 'description': ''}]

# Generate content with ethical oversight
def generate_content(
    product_name,
    product_description,
    target_audience,
    key_features,
    unique_benefits,
    platform,
    tone,
    generator_tokenizer,
    generator,
    sentiment_analyzer,
    bias_detector
):
    # Format prompt based on platform
    char_limit = 280 if platform == "Twitter" else 500
    
    # Get recent news for context
    news_data = fetch_recent_news(f"{product_name} {target_audience}")
    news_context = "\n".join([f"Recent context: {item['title']}" for item in news_data])
    
    # Create prompt
    prompt = f"""
    Product: {product_name}
    Description: {product_description}
    Target Audience: {target_audience}
    Key Features: {key_features}
    Unique Benefits: {unique_benefits}
    Tone: {tone}
    Platform: {platform}
    Character Limit: {char_limit}
    
    {news_context}
    
    Create a {platform} post that highlights the product's benefits while maintaining a {tone} tone:
    """
    
    # Generate initial content
    inputs = generator_tokenizer(prompt, return_tensors="pt", max_length=512, truncation=True)
    outputs = generator.generate(
        inputs["input_ids"],
        max_length=char_limit,
        num_return_sequences=3,
        temperature=0.7,
        top_p=0.9,
        do_sample=True,
    )
    
    generated_texts = [generator_tokenizer.decode(output, skip_special_tokens=True) for output in outputs]
    
    # Filter and analyze content
    filtered_content = []
    for text in generated_texts:
        # Clean up text
        text = text.replace(prompt, "").strip()
        
        # Skip if text is too short
        if len(text) < 10:
            continue
            
        # Check sentiment
        sentiment = sentiment_analyzer(text)[0]
        
        # Check bias
        bias = bias_detector(text)[0]
        
        # Filter based on ethical considerations
        if (
            sentiment['label'] != 'negative' and
            float(bias['score']) < 0.7 and  # Adjust threshold as needed
            len(text) <= char_limit
        ):
            filtered_content.append({
                'text': text,
                'sentiment': sentiment['label'],
                'bias_score': f"{float(bias['score']):.2f}"
            })
    
    return filtered_content

# Gradio interface
def create_interface():
    generator_tokenizer, generator, sentiment_analyzer, bias_detector = load_models()
    
    def process_input(
        product_name,
        product_description,
        target_audience,
        key_features,
        unique_benefits,
        platform,
        tone
    ):
        results = generate_content(
            product_name,
            product_description,
            target_audience,
            key_features,
            unique_benefits,
            platform,
            tone,
            generator_tokenizer,
            generator,
            sentiment_analyzer,
            bias_detector
        )
        
        if not results:
            return "No suitable content generated. Please try again with different parameters."
        
        output = ""
        for i, content in enumerate(results, 1):
            output += f"\nVersion {i}:\n"
            output += f"Content: {content['text']}\n"
            output += f"Sentiment: {content['sentiment']}\n"
            output += f"Bias Score: {content['bias_score']}\n"
            output += "-" * 50 + "\n"
        
        return output
    
    # Create the interface
    iface = gr.Interface(
        fn=process_input,
        inputs=[
            gr.Textbox(label="Product Name", placeholder="Enter product name"),
            gr.Textbox(label="Product Description", lines=3, placeholder="Brief description of your product"),
            gr.Textbox(label="Target Audience", placeholder="Who is this product for?"),
            gr.Textbox(label="Key Features", lines=2, placeholder="Main features of your product"),
            gr.Textbox(label="Unique Benefits", lines=2, placeholder="What makes your product special?"),
            gr.Radio(
                choices=["Twitter", "Instagram"],
                label="Platform",
                value="Twitter"
            ),
            gr.Textbox(label="Tone", placeholder="e.g., professional, casual, friendly"),
        ],
        outputs=gr.Textbox(label="Generated Content", lines=10),
        title="Ethimar - AI Marketing Content Generator",
        description="Generate ethical marketing content with AI-powered insights",
        theme="default"
    )
    
    return iface

# Launch the app
if __name__ == "__main__":
    iface = create_interface()
    iface.launch()