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
from custom_css import custom_css
from variables import *
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=120)
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 relevant and detailed prompt.'
},
{
"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=3000,
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.",
"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)
output={
key: value.replace('\\"', '"') if isinstance(value, str) else value
for key, value in json_output.items()
}
output['response_content']=json_output
# Clean up JSON values
return output
# 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 ""
output['response_content']=response_content
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),
'response_content':str(e)
}
def apply_prompt(self, prompt: str, model: str) -> str:
try:
messages = [
{
"role": "system",
"content": """You are a markdown formatting expert. Format your responses with proper spacing and structure following these rules:
1. Paragraph Spacing:
- Add TWO blank lines between major sections (##)
- Add ONE blank line between subsections (###)
- Add ONE blank line between paragraphs within sections
- Add ONE blank line before and after lists
- Add ONE blank line before and after code blocks
- Add ONE blank line before and after blockquotes
2. Section Formatting:
# Title
## Major Section
[blank line]
Content paragraph 1
[blank line]
Content paragraph 2
[blank line]"""
},
{
"role": "user",
"content": prompt
}
]
response = self.client.chat_completion(
model=model,
messages=messages,
max_tokens=3000,
temperature=0.8,
stream=True # Enable streaming in the API call
)
# Initialize an empty string to accumulate the response
full_response = ""
# Process the streaming response
for chunk in response:
if chunk.choices[0].delta.content is not None:
full_response += chunk.choices[0].delta.content
# Return the complete response
return full_response.replace('\n\n', '\n').strip()
except Exception as e:
return f"Error: {str(e)}"
class GradioInterface:
def __init__(self, prompt_refiner: PromptRefiner,custom_css):
self.prompt_refiner = prompt_refiner
custom_css = custom_css
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.")
gr.HTML(
"<p style='text-align: center; color:orange;'>⚠ This space is in progress, and we're actively working on it, so you might find some bugs! Please report any issues you have in the Community tab to help us make it better for all.</p>"
)
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.", "done"],
["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=True,
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]
)
# In the __init__ method of GradioInterface class:
apply_button.click(
fn=self.apply_prompts,
inputs=[prompt_text, refined_prompt, apply_model],
outputs=[original_output, refined_output],
api_name="apply_prompts" # Optional: adds API endpoint
)
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):
try:
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
except Exception as e:
return f"Error: {str(e)}", f"Error: {str(e)}"
def launch(self, share=False):
self.interface.launch(share=share)
#explanation_markdown = "".join([f"- **{key}**: {value}\n" for key, value in metaprompt_explanations.items()])
'''
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')
'''
if __name__ == '__main__':
prompt_refiner = PromptRefiner(api_token)
gradio_interface = GradioInterface(prompt_refiner,custom_css)
gradio_interface.launch(share=True)