import gradio as gr
import os

# HF Serverless Inference API - free tier, built into Spaces
from huggingface_hub import InferenceClient

# Free model accessible via HF serverless
MODEL = "Qwen/Qwen2.5-7B-Instruct"

with open("SOUL.md") as f:
    SOUL_TEMPLATE = f.read()

def compare(prompt):
    if not prompt.strip():
        return "Enter a prompt.", "", ""
    
    try:
        client = InferenceClient(token=os.environ.get("HF_TOKEN", None))
        
        # Raw response - no system prompt
        raw = client.chat.completions.create(
            model=MODEL,
            messages=[{"role": "user", "content": prompt}],
            max_tokens=500,
            temperature=0.7,
        )
        raw_text = raw.choices[0].message.content
        
        # With SOUL injected
        soul = client.chat.completions.create(
            model=MODEL,
            messages=[
                {"role": "system", "content": SOUL_TEMPLATE},
                {"role": "user", "content": prompt}
            ],
            max_tokens=500,
            temperature=0.7,
        )
        soul_text = soul.choices[0].message.content
        
        cost = "Model: Qwen2.5-7B (free) | Same model, same cost — different SOUL"
        return raw_text, soul_text, cost
        
    except Exception as e:
        return f"Error: {str(e)[:200]}", "Please try again later.", ""

with gr.Blocks(title="Poor Mans Opus Before After", theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
    # 🦾 Poor Man's Opus — Live Comparison
    
    **Our product is the SOUL, not the model.**
    
    Left: raw model output. Right: same model + Poor Man's Opus `::GENE{}` behavioral DNA.
    Running on a **free** model right now. The SOUL works on ANY model.
    """)
    
    gr.Markdown("---")
    
    prompt = gr.Textbox(
        label="Ask anything",
        placeholder="e.g. What are the three biggest mistakes founders make raising Series A?",
        lines=3
    )
    
    btn = gr.Button("Compare", variant="primary", size="lg")
    
    with gr.Row():
        with gr.Column():
            gr.Markdown("### 🔴 Raw (Default Behavior)")
            raw_out = gr.Textbox(label="", lines=18)
        with gr.Column():
            gr.Markdown("### 🟢 Same Model + Poor Man's Opus SOUL")
            soul_out = gr.Textbox(label="", lines=18)
    
    stats = gr.Textbox(label="Details", interactive=False)
    
    btn.click(compare, inputs=prompt, outputs=[raw_out, soul_out, stats])
    
    # Load default example
    demo.load(
        lambda: compare("Explain how LLMs work to a 10-year-old"),
        outputs=[raw_out, soul_out, stats]
    )
    
    gr.Markdown("""
    ---
    ### The SOUL works on any model
    
    This demo runs on **Qwen2.5-7B** — a free, open-source model. The same `::GENE{}` DNA changes its behavior immediately.
    
    **For the best results, use a premium reasoning model:**
    
    | Model | Cost (input/output per M tokens) | Verdict |
    |-------|------|---------|
    | DeepSeek V4 Pro | $1.74 / $3.48 | 🏆 Best ROI — Opus-level output at 3% cost |
    | DeepSeek V4 Flash | $0.14 / $0.28 | Fast & cheap |
    | Claude Opus 4.6 | $15.00 / $75.00 | Baseline quality (what we compare to) |
    
    **One command:**
    ```bash
    openclaw skills install poor-mans-opus
    ```
    
    [📦 GitHub](https://github.com/mtmpss/poor-mans-opus) | [🛒 ClawHub](https://clawhub.ai/mtmpss/poor-mans-opus) | [🤗 HF Model](https://huggingface.co/ilanguage/poor-mans-opus)
    """)

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