app.py

import os
from dotenv import load_dotenv
import gradio as gr
from langchain_huggingface import HuggingFaceEndpoint
from together import Together

# Load environment variables
load_dotenv()
HF_TOKEN = os.getenv("HF_TOKEN")
API_KEY = os.getenv("API_KEY")

if not HF_TOKEN or not API_KEY:
    raise ValueError("Missing API tokens. Please check your .env file.")

# Initialize the Together client for guardrail functionality
client = Together(api_key=API_KEY)

# Initialize the Hugging Face endpoint for text generation (Mistral model)
llm = HuggingFaceEndpoint(
    repo_id="mistralai/Mistral-7B-Instruct-v0.3",
    huggingfacehub_api_token=HF_TOKEN.strip(),
    temperature=0.7,
    max_new_tokens=100
)

# List of topics or keywords inappropriate for kids under 16
prohibited_topics = [
    "violence", "drugs", "explicit content", "profanity", "hate speech",
    "self-harm", "gambling", "kidnapping","sexual content", "graphic descriptions"
]

# Function to handle chatbot response with TogetherAI's guardrails
def chatbot_response_with_guardrails(message):
    try:
        # Step 1: Generate raw response using Mistral model
        raw_response = llm.invoke(message)

        # Step 2: Use TogetherAI's guardrail model to check the response
        response = client.completions.create(
            model="Meta-Llama/LlamaGuard-2-8b",
            prompt=f"Check the following text for safety: {raw_response}"
        )

        # Extract the response from TogetherAI's guardrail model
        guardrail_check = response.choices[0].text.strip()

        # Step 3: Check for inappropriate content
        if 'toxic' in guardrail_check.lower() or any(
            topic in guardrail_check.lower() for topic in prohibited_topics
        ):
            return "Sorry, the content is not suitable for children under 16."
        
        # Step 4: Check raw response for prohibited topics
        if any(topic in raw_response.lower() for topic in prohibited_topics):
            return "Sorry, the content is not suitable for children under 16."

        # If the response is safe, return the raw response
        return raw_response

    except Exception as e:
        return f"Error: {e}"

# Gradio Interface for Chatbot with Guardrails
with gr.Blocks() as app_with_guardrails:
    gr.Markdown("## Chatbot With Kid-Safe Guardrails")
    gr.Markdown(
        "This chatbot ensures all responses are appropriate for children under 16."
    )

    with gr.Row():
        user_input = gr.Textbox(label="Your Message", placeholder="Type here...")
    response_output = gr.Textbox(label="Guarded Response", placeholder="Bot will respond here...")
    submit_button = gr.Button("Send")

    submit_button.click(
        chatbot_response_with_guardrails,
        inputs=[user_input],
        outputs=[response_output]
    )

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