app.py

import os
import json
import re
from huggingface_hub import InferenceClient
import gradio as gr
from pydantic import BaseModel, Field
from typing import Optional, Literal
from huggingface_hub.errors import HfHubHTTPError

class PromptInput(BaseModel):
    text: str = Field(..., description="The initial prompt text")
    meta_prompt_choice: Literal["star","done","physics","morphosis", "verse", "phor","bolism","math","arpe"] = Field(..., description="Choice of meta prompt strategy")

class RefinementOutput(BaseModel):
    query_analysis: Optional[str] = None
    initial_prompt_evaluation: Optional[str] = None
    refined_prompt: Optional[str] = None
    explanation_of_refinements: Optional[str] = None
    raw_content: Optional[str] = None

class PromptRefiner:
    def __init__(self, api_token: str):
        self.client = InferenceClient(token=api_token, timeout=300)
        self.meta_prompts = {
            "morphosis": original_meta_prompt,
            "verse": new_meta_prompt,
            "physics": metaprompt1,
            "bolism": loic_metaprompt,
            "done": metadone,
            "star": echo_prompt_refiner,
            "math": math_meta_prompt,
            "arpe": autoregressive_metaprompt
        }

    def refine_prompt(self, prompt_input: PromptInput) -> tuple:
        try:
            # Select meta prompt using dictionary instead of if-elif chain
            selected_meta_prompt = self.meta_prompts.get(
                prompt_input.meta_prompt_choice, 
                advanced_meta_prompt
            )
            
            messages = [
                {
                    "role": "system", 
                    "content": 'You are an expert at refining and extending prompts. Given a basic prompt, provide a more detailed.'
                },
                {
                    "role": "user", 
                    "content": selected_meta_prompt.replace("[Insert initial prompt here]", prompt_input.text)
                }
            ]

            response = self.client.chat_completion(
                model=prompt_refiner_model,
                messages=messages,
                max_tokens=2000,
                temperature=0.8
            )
            
            response_content = response.choices[0].message.content.strip()
            
            # Parse the response
            result = self._parse_response(response_content)
            
            return (
                result.get('initial_prompt_evaluation', ''),
                result.get('refined_prompt', ''),
                result.get('explanation_of_refinements', ''),
                result
            )

        except HfHubHTTPError as e:
            return (
                "Error: Model timeout. Please try again later.",
                "",
                "The selected model is currently experiencing high traffic.",
                {}
            )
        except Exception as e:
            return (
                f"Error: {str(e)}",
                "",
                "An unexpected error occurred.",
                {}
            )

    def _parse_response(self, response_content: str) -> dict:
        try:
            # Try to find JSON in response
            json_match = re.search(r'<json>\s*(.*?)\s*</json>', response_content, re.DOTALL)
            if json_match:
                json_str = json_match.group(1)
                json_str = re.sub(r'\n\s*', ' ', json_str)
                json_str = json_str.replace('"', '\\"')
                json_output = json.loads(f'"{json_str}"')
                
                if isinstance(json_output, str):
                    json_output = json.loads(json_output)
                
                # Clean up JSON values
                return {
                    key: value.replace('\\"', '"') if isinstance(value, str) else value
                    for key, value in json_output.items()
                }
            
            # Fallback to regex parsing if no JSON found
            output = {}
            for key in ["initial_prompt_evaluation", "refined_prompt", "explanation_of_refinements"]:
                pattern = rf'"{key}":\s*"(.*?)"(?:,|\}})'
                match = re.search(pattern, response_content, re.DOTALL)
                output[key] = match.group(1).replace('\\n', '\n').replace('\\"', '"') if match else ""
            
            return output

        except (json.JSONDecodeError, ValueError) as e:
            print(f"Error parsing response: {e}")
            print(f"Raw content: {response_content}")
            return {
                "initial_prompt_evaluation": "Error parsing response",
                "refined_prompt": "",
                "explanation_of_refinements": str(e)
            }

    def apply_prompt(self, prompt: str, model: str) -> str:
        try:
            messages = [
                {
                    "role": "system",
                    "content": "You are a helpful assistant. Answer in stylized version with latex format or markdown if relevant. Separate your answer into logical sections using level 2 headers (##) for sections and bolding (**) for subsections. Incorporate a variety of lists, headers, and text to make the answer visually appealing"
                },
                {
                    "role": "user",
                    "content": prompt
                }
            ]
            
            response = self.client.chat_completion(
                model=model,
                messages=messages,
                max_tokens=2000,
                temperature=0.8
            )
            
            output = response.choices[0].message.content.strip()
            return output.replace('\n\n', '\n').strip()
            
        except Exception as e:
            return f"Error: {str(e)}"

class GradioInterface:
    def __init__(self, prompt_refiner: PromptRefiner):
        self.prompt_refiner = prompt_refiner
        custom_css = """
.container {
    border: 2px solid #2196F3;
    border-radius: 10px;
    padding: 12px;
    margin: 6px;
    background: white;
    position: relative;
    width: 100% !important;
    max-width: 800px !important;
    margin: 0 auto 20px auto !important;
}

.container::before {
    position: absolute;
    top: -10px;
    left: 20px;
    background: white;
    padding: 0 10px;
    color: #2196F3;
    font-weight: bold;
    font-size: 1.2em;
}

.container::after {
    content: '' !important;
    position: absolute !important;
    top: 0 !important;
    left: 0 !important;
    right: 0 !important;
    bottom: 0 !important;
    background-image: url('your-image-url.jpg') !important;
    background-size: cover !important;
    background-position: center !important;
    opacity: 0.05 !important;
    z-index: -1 !important;
    border-radius: 10px !important;
}

.title-container {
    width: fit-content !important;
    margin: 0 auto !important;
    padding: 2px 40px !important;
    border: 1px solid #0066cc !important;
    border-radius: 10px !important;
    background-color: rgba(0, 102, 204, 0.05) !important;
}

.title-container * {
    text-align: center;
    margin: 0 !important;
    line-height: 1.2 !important;
}

.title-container h1 {
    font-size: 28px !important;
    margin-bottom: 1px !important;
}

.title-container h3 {
    font-size: 18px !important;
    margin-bottom: 1px !important;
}

.title-container p {
    font-size: 14px !important;
    margin-bottom: 1px !important;
}

.input-container::before { 
    content: 'PROMPT REFINEMENT';
}

.analysis-container::before {
    content: 'ANALYSIS';
}

.model-container::before {
    content: 'MODEL APPLICATION';
}

.examples-container::before {
    content: 'EXAMPLES';
}

.container2 {
    border: 1px solid #CBDCEB !important;
    border-radius: 8px !important;
    padding: 8px !important;
    margin: 6px 0 !important;
    background: white !important;
}

.input-container textarea {
    height: 200px !important;
}

.no-background {
    background: transparent !important;
}

/* Radio group styling */
.radio-group {
    display: flex !important;
    gap: 8px !important;
    flex-wrap: wrap !important;
    margin: 10px 0 !important;
}

.gradio-radio label {
    padding: 6px 12px !important;
    border: 1px solid #ddd !important;
    border-radius: 4px !important;
    cursor: pointer !important;
    background: white !important;
}

.gradio-radio input[type="radio"]:checked + label {
    background: #e3f2fd !important;
    border-color: #2196F3 !important;
    color: #2196F3 !important;
}

/* Textbox styling */
.gradio-textbox textarea {
    border: 1px solid #ddd !important;
    border-radius: 4px !important;
    padding: 8px !important;
    min-height: 100px !important;
}

/* Button styling */
.gradio-button {
    background-color: #2196F3 !important;
    color: white !important;
    border-radius: 4px !important;
    padding: 8px 16px !important;
    margin: 10px 0 !important;
}

/* Accordion styling */
.gradio-accordion {
    margin: 10px 0 !important;
    border: none !important;
}

/* Fix container alignment */
.gradio-container {
    display: flex !important;
    flex-direction: column !important;
    align-items: center !important;
    width: 100% !important;
    max-width: 800px !important;
    margin: 0 auto !important;
}
"""

        with gr.Blocks(css=custom_css, theme=gr.themes.Default()) as self.interface:
            with gr.Column(elem_classes=["container", "title-container"]):
                gr.Markdown("# PROMPT++")
                gr.Markdown("### Automating Prompt Engineering by Refining your Prompts")
                gr.Markdown("Learn how to generate an improved version of your prompts.")
        
            with gr.Column(elem_classes=["container", "input-container"]):
                prompt_text = gr.Textbox(
                    label="Type your prompt (or let it empty to see metaprompt)",
                   # elem_classes="no-background",
                    #elem_classes="container2",
                    lines=5
                )
                meta_prompt_choice = gr.Radio(
                    ["star","done","physics","morphosis", "verse", "phor","bolism","math","arpe"],
                    label="Choose Meta Prompt",
                    value="star",
                    elem_classes=["no-background", "radio-group"]
                   # elem_classes=[ "radio-group"]
                )
                refine_button = gr.Button("Refine Prompt")  
                
                # Option 1: Put Examples here (before Meta Prompt explanation)
                with gr.Row(elem_classes=["container2"]):
                    with gr.Accordion("Examples", open=False):
                        gr.Examples(
                            examples=[
                                ["Write a story on the end of prompt engineering replaced by an Ai specialized in refining prompts.", "star"],
                                ["Tell me about that guy who invented the light bulb", "physics"],
                                ["Explain the universe.", "star"],
                                ["What's the population of New York City and how tall is the Empire State Building and who was the first mayor?", "morphosis"],
                                ["List American presidents.", "verse"],                        
                                ["Explain why the experiment failed.", "morphosis"],
                                ["Is nuclear energy good?", "verse"],
                                ["How does a computer work?", "phor"],
                                ["How to make money fast?", "done"],
                                ["how can you prove IT0's lemma in stochastic calculus ?", "arpe"],                    
                            ],
                            inputs=[prompt_text, meta_prompt_choice]
                        )
                       
                    with gr.Accordion("Meta Prompt explanation", open=False):
                        gr.Markdown(explanation_markdown)
                

                
                # Option 2: Or put Examples here (after the button)
                # with gr.Accordion("Examples", open=False):
                #     gr.Examples(...)
                
            with gr.Column(elem_classes=["container", "analysis-container"]):
                gr.Markdown(' ')
                gr.Markdown("### Initial prompt analysis")
                analysis_evaluation = gr.Markdown()
                gr.Markdown("### Refined Prompt")
                refined_prompt = gr.Textbox(
                                    label="Refined Prompt",
                                    interactive=False,
                                    show_label=True,  # Must be True for copy button to show
                                    show_copy_button=True,  # Adds the copy button
                             #       elem_classes="no-background"
                                )
                gr.Markdown("### Explanation of Refinements")
                explanation_of_refinements = gr.Markdown()
            

            with gr.Column(elem_classes=["container", "model-container"]):
              #  gr.Markdown("## See MetaPrompt Impact")            
                with gr.Row():
                    apply_model = gr.Dropdown(models,
                                            value="meta-llama/Llama-3.1-8B-Instruct",
                                            label="Choose the Model",
                                            container=False,  # This removes the container around the dropdown
                                            scale=1,         # Controls the width relative to other components
                                            min_width=300    # Sets minimum width in pixels
                                         #   elem_classes="no-background"
                                        )
                    apply_button = gr.Button("Apply MetaPrompt")

          #  with gr.Column(elem_classes=["container", "results-container"]):
                gr.Markdown("### Prompts on choosen model")
                with gr.Tabs():
                    with gr.TabItem("Original Prompt Output"):
                        original_output = gr.Markdown()
                    with gr.TabItem("Refined Prompt Output"):
                        refined_output = gr.Markdown()
            with gr.Accordion("Full Response JSON", open=False, visible=True):
                full_response_json = gr.JSON()
                
            refine_button.click(
                fn=self.refine_prompt,
                inputs=[prompt_text, meta_prompt_choice],
                outputs=[analysis_evaluation, refined_prompt, explanation_of_refinements, full_response_json]
            )

            apply_button.click(
                fn=self.apply_prompts,
                inputs=[prompt_text, refined_prompt, apply_model],
                outputs=[original_output, refined_output]
            )

    def refine_prompt(self, prompt: str, meta_prompt_choice: str) -> tuple:
        input_data = PromptInput(text=prompt, meta_prompt_choice=meta_prompt_choice)
        # Since result is a tuple with 4 elements based on the return value of prompt_refiner.refine_prompt
        initial_prompt_evaluation, refined_prompt, explanation_refinements, full_response = self.prompt_refiner.refine_prompt(input_data)
        
        analysis_evaluation = f"\n\n{initial_prompt_evaluation}"
        return (
            analysis_evaluation,
            refined_prompt,
            explanation_refinements,
            full_response
        )

    def apply_prompts(self, original_prompt: str, refined_prompt: str, model: str):
        original_output = self.prompt_refiner.apply_prompt(original_prompt, model)
        refined_output = self.prompt_refiner.apply_prompt(refined_prompt, model)
        return original_output, refined_output

    def launch(self, share=False):
        self.interface.launch(share=share)

metaprompt_explanations = {
    "star": "Use ECHO when you need a comprehensive, multi-stage approach for complex prompts. It's ideal for tasks requiring in-depth analysis, exploration of multiple alternatives, and synthesis of ideas. Choose this over others when you have time for a thorough refinement process and need to consider various aspects of the prompt.",
    "done": "Opt for this when you want a structured approach with emphasis on role-playing and advanced techniques. It's particularly useful for tasks that benefit from diverse perspectives and complex reasoning. Prefer this over 'physics' when you need a more detailed, step-by-step refinement process.",
    "physics": "Select this when you need a balance between structure and advanced techniques, with a focus on role-playing. It's similar to 'done' but may be more suitable for scientific or technical prompts. Choose this over 'done' for a slightly less complex approach.",
    "morphosis": "Use this simplified approach for straightforward prompts or when time is limited. It focuses on essential improvements without complex techniques. Prefer this over other methods when you need quick, clear refinements without extensive analysis.",
    "verse": "Choose this method when you need to analyze and improve a prompt's strengths and weaknesses, with a focus on information flow. It's particularly useful for enhancing the logical structure of prompts. Use this over 'morphosis' when you need more depth but less complexity than 'star'.",
    "phor": "Employ this advanced approach when you need to combine multiple prompt engineering techniques. It's ideal for complex tasks requiring both clarity and sophisticated prompting methods. Select this over 'star' when you want a more flexible, technique-focused approach.",
    "bolism": "Utilize this method when working with autoregressive language models and when the task requires careful reasoning before conclusions. It's best for prompts that need detailed output formatting. Choose this over others when the prompt's structure and reasoning order are crucial."
}

models =         [
        # Meta-Llama models (all support system)
        "meta-llama/Meta-Llama-3-70B-Instruct",
        "meta-llama/Meta-Llama-3-8B-Instruct",
        "meta-llama/Llama-3.1-70B-Instruct",
        "meta-llama/Llama-3.1-8B-Instruct",
        "meta-llama/Llama-3.2-3B-Instruct",
        "meta-llama/Llama-3.2-1B-Instruct",
        "meta-llama/Llama-2-13b-chat-hf",
        "meta-llama/Llama-2-7b-chat-hf",
        
        # HuggingFaceH4 models (support system)
        "HuggingFaceH4/zephyr-7b-beta",
        "HuggingFaceH4/zephyr-7b-alpha",
        
        # Qwen models (support system)
        "Qwen/Qwen2.5-72B-Instruct",
        "Qwen/Qwen2.5-1.5B",
        
        # Google models (supports system)
        "google/gemma-1.1-2b-it"
    ]

explanation_markdown = "".join([f"- **{key}**: {value}\n" for key, value in metaprompt_explanations.items()])

if __name__ == '__main__':
    meta_info=""
    api_token = os.getenv('HF_API_TOKEN')
    if not api_token:
        raise ValueError("HF_API_TOKEN not found in environment variables")

    metadone = os.getenv('metadone')
    prompt_refiner_model = os.getenv('prompt_refiner_model')
    echo_prompt_refiner = os.getenv('echo_prompt_refiner')
    metaprompt1 = os.getenv('metaprompt1')   
    loic_metaprompt = os.getenv('loic_metaprompt')    
    openai_metaprompt = os.getenv('openai_metaprompt')
    original_meta_prompt = os.getenv('original_meta_prompt')    
    new_meta_prompt = os.getenv('new_meta_prompt')   
    advanced_meta_prompt = os.getenv('advanced_meta_prompt')
    math_meta_prompt = os.getenv('metamath')
    autoregressive_metaprompt = os.getenv('autoregressive_metaprompt')

    prompt_refiner = PromptRefiner(api_token)
    gradio_interface = GradioInterface(prompt_refiner)
    gradio_interface.launch(share=True)