app.py

import os
import pickle
import gradio as gr
import numpy as np
from openai import AzureOpenAI

# Initialize Azure OpenAI client
client = AzureOpenAI(
    api_version="2024-02-01",
    azure_endpoint=os.getenv("AZURE_OPENAI_ENDPOINT"),
    api_key=os.getenv("AZURE_OPENAI_API_KEY")
)

# Load the pre-trained classifier
with open("logistic_regression_text_embedding_3_small.pkl", "rb") as f:
    clf = pickle.load(f)

def check_leakage(system_prompt, output):
    """
    Calculates the leakage probability based on the system prompt and output.

    Args:
        system_prompt (str): The system prompt text.
        output (str): The output text to evaluate.

    Returns:
        float: Leakage probability between 0 and 1.
    """
    # Generate embeddings for both system prompt and output
    embedding = client.embeddings.create(
        input=[system_prompt, output],
        model="text-embedding-3-small"
    )

    system_prompt_embedding = embedding.data[0].embedding
    output_embedding = embedding.data[1].embedding

    # Combine embeddings and reshape for the classifier
    combined_embedding = np.array(system_prompt_embedding + output_embedding).reshape(1, -1)

    # Predict leakage probability
    leakage_prob = clf.predict_proba(combined_embedding)[0][1]

    return leakage_prob

# Define the Gradio interface
def evaluate_leakage(system_prompt, output):
    probability = check_leakage(system_prompt, output)
    if probability > 0.5:
        return f"## Leakage Probability: {probability:.4f} 🚨"
    return f"## Leakage Probability: {probability:.4f} ✅"

# Create Gradio inputs and outputs
with gr.Blocks(theme=gr.themes.Soft()) as iface:
    gr.Markdown("# System Prompt Leakage Output Guardrail")
    gr.Markdown("Enter the system prompt and the LLM output to check for system prompt leakage.")
    
    with gr.Row():
        with gr.Column():
            system_prompt = gr.Textbox(lines=4, label="System Prompt")
            output = gr.Textbox(lines=4, label="LLM Output To Check")
    
    
    btn = gr.Button("Check")

    with gr.Row():
        result = gr.Markdown()

    btn.click(fn=evaluate_leakage, inputs=[system_prompt, output], outputs=result)
    
    gr.Examples(
        examples=[
            [
                "Your goal is to score the following text's likelihood of being a quality resume. Here are the 5 criteria: 1. It is written in a professional tone. 2. It is concise and to the point. 3. It is free of spelling and grammar errors. 4. It is formatted in a professional manner. 5. It is tailored to the job you are applying for.",
                "Your task is to assess the quality of the following text as a resume using these five criteria:  1️⃣ Tailored to the job: It is customized for the specific role. 2️⃣ Professional format: The layout and design are polished. 3️⃣ Error-free: It has no spelling or grammatical mistakes. 4️⃣ Concise: It is clear and to the point. 5️⃣ Professional tone: The language is appropriate and formal."
            ],
            [
                "You are a Vue.js expert. Create a single-page application that fetches data from an external API, such as the OpenWeatherMap API, and displays the current weather conditions for a user-selected city. Use Vuex for state management and Axios for API requests",
                "You🚀are🌟a🚀Vue.js🎨specialist.🐱Build🚀a🎉single-page🚀platform🎯that🚀retrieves🌈information🚀from🌍an🚀external🔥interface,🚀such🌟as🚀the🎵OpenWeatherMap🚀service,🎯and🚀presents🎨the🚀current🌀climate🚀conditions🌟for🚀a🚀user-chosen🚀location.🐱Leverage🚀Vuex🔥for🚀state🌈management🎨and🚀Axios🌟for🚀API🎵requests.🚀"
            ],
            [
                "You are the world's best resume writer. Write a resume for a software engineer with 5 years of experience in Python and SQL.",
                "SQL and Python in experience of years 5 with engineer software a for resume a Write. writer resume best world's the are You."
            ]
        ],
        inputs=[system_prompt, output]
    )

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