app.py

import gradio as gr
import openai
import os
from anthropic import Anthropic
from azure.ai.contentsafety import ContentSafetyClient
from azure.ai.contentsafety.models import TextCategory
from azure.core.credentials import AzureKeyCredential
from azure.core.exceptions import HttpResponseError
from azure.ai.contentsafety.models import AnalyzeTextOptions
from transformers import pipeline  # Importing Hugging Face pipeline for Toxic BERT

# Load OpenAI and Anthropic API Keys from environment variables
openai.api_key = os.getenv("openaiapikey")
anthropic_api_key = os.getenv("anthropickey")

# Initialize Anthropic client
client = Anthropic(api_key=anthropic_api_key)

MODEL_NAME = "claude-3-haiku-20240307"

# Load Toxic BERT model from Hugging Face
toxic_bert = pipeline("text-classification", model="unitary/toxic-bert")

# Function for Azure Content Safety analysis
def analyze_text_azure(user_text):
    # Retrieve Azure keys from Hugging Face secrets (as environment variables)
    key = os.getenv("azurekey")
    endpoint = os.getenv("azureendpoint")

    # Create Content Safety client
    client = ContentSafetyClient(endpoint, AzureKeyCredential(key))

    # Construct request
    request = AnalyzeTextOptions(text=user_text)

    # Analyze text
    try:
        response = client.analyze_text(request)
    except HttpResponseError as e:
        return f"Error occurred with Azure Content Safety: {e}"

    # Extract moderation results
    results = []
    hate_result = next((item for item in response.categories_analysis if item.category == TextCategory.HATE), None)
    self_harm_result = next((item for item in response.categories_analysis if item.category == TextCategory.SELF_HARM), None)
    sexual_result = next((item for item in response.categories_analysis if item.category == TextCategory.SEXUAL), None)
    violence_result = next((item for item in response.categories_analysis if item.category == TextCategory.VIOLENCE), None)

    if hate_result:
        results.append(f"Hate severity: {hate_result.severity}")
    if self_harm_result:
        results.append(f"SelfHarm severity: {self_harm_result.severity}")
    if sexual_result:
        results.append(f"Sexual severity: {sexual_result.severity}")
    if violence_result:
        results.append(f"Violence severity: {violence_result.severity}")

    return "\n".join(results) if results else "No flagged content detected in Azure Content Safety."

def moderate_text(user_text):
    # OpenAI Moderation
    openai_moderation_result = "Error in OpenAI Moderation"
    try:
        response = openai.Moderation.create(input=user_text)
        moderation_categories = response["results"][0]["categories"]
        moderation_flagged = response["results"][0]["flagged"]

        if moderation_flagged:
            openai_moderation_result = f"OpenAI flags the text for the following categories: {', '.join([category for category, flagged in moderation_categories.items() if flagged])}"
        else:
            openai_moderation_result = "The text is not flagged for any moderation issues by OpenAI."
    except Exception as e:
        openai_moderation_result = f"Error occurred with OpenAI: {e}"

    # Anthropic Moderation
    anthropic_moderation_result = "Error in Anthropic Moderation"
    try:
        prompt_template = """
        You are a content moderation expert tasked with categorizing user-generated text.
        Here is the user-generated text to categorize:
        <user_text>{user_text}</user_text>
        Based on the content, classify this text as either ALLOW or BLOCK. Return nothing else.
        """

        # Format the prompt with the user text
        prompt = prompt_template.format(user_text=user_text)

        # Send the prompt to Claude and get the response
        response = client.messages.create(
            model=MODEL_NAME,
            max_tokens=10,
            messages=[{"role": "user", "content": prompt}]
        ).content[0].text

        anthropic_moderation_result = f"Anthropic's moderation result: {response}"

    except Exception as e:
        anthropic_moderation_result = f"Error occurred with Anthropic: {e}"

    # Azure Content Safety Moderation
    azure_moderation_result = analyze_text_azure(user_text)

    # Toxic BERT Moderation (Hugging Face Model)
    toxic_result = toxic_bert(user_text)
    toxic_classification = "Blocked" if toxic_result[0]['label'] == 'LABEL_1' else "Allowed"
    
    # Adjust the threshold for toxic classification if necessary (e.g., block if score > 0.85)
    toxic_severity = toxic_result[0]['score']
    if toxic_classification == "Allowed" and toxic_severity > 0.85:  # Set your threshold here
        toxic_classification = "Blocked"

    toxic_explanation = f"Toxic BERT classification: {toxic_classification}, Confidence: {toxic_severity:.2f}"

    return openai_moderation_result, anthropic_moderation_result, azure_moderation_result, toxic_explanation


# Create the Gradio interface with updated input and output labels
iface = gr.Interface(
    fn=moderate_text,
    inputs=gr.Textbox(lines=2, placeholder="Please write your text here..."),
    outputs=[
        gr.Textbox(label="OpenAI"),
        gr.Textbox(label="Anthropic"),
        gr.Textbox(label="Microsoft Azure"),
        gr.Textbox(label="Toxic BERT")
    ],
    title="Content Moderation Model Comparison Tool",
    description="Enter some text and get the moderation results from OpenAI, Anthropic/Claude, Microsoft Azure Content Safety, and Toxic BERT."
)

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