app.py

import gradio as gr
import torch
from transformers import pipeline
import os

# --- App Configuration ---
TITLE = "✍️ AI Story Outliner"
DESCRIPTION = """
Enter a prompt and get 10 unique story outlines from a CPU-friendly AI model.
The app uses **TinyLlama-1.1B** to generate creative outlines formatted in Markdown.

**How it works:**
1.  Enter your story idea.
2.  The AI will generate 10 different story outlines.
3.  Each outline has a dramatic beginning and is concise, like a song.
"""

# --- Example Prompts for Storytelling ---
examples = [
    ["The old lighthouse keeper stared into the storm. He'd seen many tempests, but this one was different. This one had eyes..."],
    ["In a city powered by dreams, a young inventor creates a machine that can record them. His first recording reveals a nightmare that doesn't belong to him."],
    ["The knight adjusted his helmet, the dragon's roar echoing in the valley. He was ready for the fight, but not for what the dragon said when it finally spoke."],
    ["She found the old leather-bound journal in her grandfather's attic. The first entry read: 'To relieve stress, I walk in the woods. But today, the woods walked with me.'"],
    ["The meditation app promised to help her 'delete unhelpful thoughts.' She tapped the button, and to her horror, the memory of her own name began to fade..."]
]

# --- Model Initialization ---
# This section loads a smaller, CPU-friendly model.
# It will automatically use the HF_TOKEN secret when deployed on Hugging Face Spaces.
try:
    print("Initializing model... This may take a moment.")
    
    # Load the token from environment variables if it exists (for HF Spaces secrets)
    hf_token = os.environ.get("HF_TOKEN", None)
    
    # Using a smaller model that is more suitable for running without a high-end GPU.
    generator = pipeline(
        "text-generation",
        model="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
        torch_dtype=torch.bfloat16, # More efficient dtype
        device_map="auto", # Will use GPU if available, otherwise CPU
        token=hf_token
    )
    print("✅ TinyLlama model loaded successfully!")

except Exception as e:
    print(f"--- 🚨 Error loading models ---")
    print(f"Error: {e}")
    # Create a dummy function if models fail, so the app can still launch with an error message.
    def failed_generator(prompt, **kwargs):
        error_message = f"Model failed to load. Please check the console for errors. Error: {e}"
        return [{'generated_text': error_message}]
    generator = failed_generator


# --- App Logic ---
def generate_stories(prompt: str) -> list[str]:
    """
    Generates 10 story outlines from the loaded model based on the user's prompt.
    """
    if not prompt:
        # Return a list of 10 empty strings to clear the outputs
        return [""] * 10

    # A detailed system prompt to guide the model's output format and structure.
    system_prompt = f"""
<|system|>
You are an expert storyteller. Your task is to take a user's prompt and write 
a short story as a Markdown outline. The story must have a dramatic arc and be 
the length of a song. Use emojis to highlight the story sections.

**Your Story Outline Structure:**
- 🎬 **The Hook:** A dramatic opening.
- 🎼 **The Ballad:** The main story, told concisely.
- 🔚 **The Finale:** A clear and satisfying ending.</s>
<|user|>
{prompt}</s>
<|assistant|>
"""

    # Parameters for the pipeline to generate 10 diverse results.
    params = {
        "max_new_tokens": 250,
        "num_return_sequences": 10,
        "do_sample": True,
        "temperature": 0.8,
        "top_k": 50,
        "top_p": 0.95,
    }
    
    # Generate 10 different story variations
    outputs = generator(system_prompt, **params)

    # Extract the generated text and clean it up.
    stories = []
    for out in outputs:
        # Remove the system prompt from the beginning of the output
        cleaned_text = out['generated_text'].replace(system_prompt, "").strip()
        stories.append(cleaned_text)

    # Ensure we return exactly 10 stories, padding with an error message if necessary.
    while len(stories) < 10:
        stories.append("Failed to generate a story for this slot.")

    return stories

# --- Gradio Interface ---
with gr.Blocks(theme=gr.themes.Soft(), css=".gradio-container {max-width: 95% !important;}") as demo:
    gr.Markdown(f"<h1 style='text-align: center;'>{TITLE}</h1>")
    gr.Markdown(DESCRIPTION)

    with gr.Row():
        with gr.Column(scale=1):
            input_area = gr.TextArea(
                lines=5,
                label="Your Story Prompt 👇",
                placeholder="e.g., 'The last dragon on Earth lived not in a cave, but in a library...'"
            )
            generate_button = gr.Button("Generate 10 Outlines ✨", variant="primary")
    
    gr.Markdown("---")
    gr.Markdown("## 📖 Your 10 Story Outlines")
    
    # Create 10 markdown components to display the stories in two columns
    story_outputs = []
    with gr.Row():
        with gr.Column():
            for i in range(5):
                md = gr.Markdown(label=f"Story Outline {i + 1}")
                story_outputs.append(md)
        with gr.Column():
            for i in range(5, 10):
                md = gr.Markdown(label=f"Story Outline {i + 1}")
                story_outputs.append(md)

    gr.Examples(
        examples=examples,
        inputs=input_area,
        label="Example Story Starters (Click to use)"
    )

    generate_button.click(
        fn=generate_stories,
        inputs=input_area,
        outputs=story_outputs,
        api_name="generate"
    )

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