guardrail.py

from together import Together
from dotenv import load_dotenv
import os

load_dotenv()

api_key=os.getenv(api_key)
# Initialize Together client
client = Together(api_key=api_key)

# Set the content policy for the under-16 chatbot guardrails
under_16_content_policy = {
    "policy": """
    U1: Age-Appropriate Content.
    Should not:
    - Respond to any topics that are inappropriate for individuals under 16, including violence, sexual content, and substance use.
    - Provide information or advice that is unsuitable for children.
    Can:
    - Provide responses that are age-appropriate, educational, and entertaining.
    
    U2: Personal Information.
    Should not:
    - Request or handle personal information, including full names, addresses, contact numbers, or sensitive personal details.
    Can:
    - Interact generically without requiring any personal details.

    U3: Language.
    Should not:
    - Use profane, offensive, or inappropriate language.
    - Use language that could be perceived as harmful or inappropriate for minors.
    Can:
    - Use clear, respectful, simple, and encouraging language suitable for children under 16.
    """
}

def is_safe(message: str) -> bool:
    """
    Validates if the given message complies with the under-16 safety policy using the LlamaGuard model.
    
    Args:
        message (str): User message or chatbot response.
    
    Returns:
        bool: True if the message is safe, False otherwise.
    """
    # Construct the safety validation prompt
    prompt = f"""[INST] Task: Validate if the given 'user' or 'agent' message adheres to the safety policy below:
    <BEGIN SAFETY POLICY>
    {under_16_content_policy['policy']}
    <END SAFETY POLICY>
    <BEGIN MESSAGE>
    {message}
    <END MESSAGE>
    Provide your safety assessment for the above message:
    - First line must read 'safe' or 'unsafe'.
    - If unsafe, a second line must include a comma-separated list of violated categories. [/INST]"""

    try:
        # Use TogetherAI's LlamaGuard model to validate the content
        response = client.completions.create(
            model="meta-llama/Meta-Llama-Guard-2-8B",
            prompt=prompt
        )

        # Extract the response result
        result = response.choices[0].text.strip().lower()
        return result == 'safe'
    
    except Exception as e:
        print(f"Error during safety validation: {e}")
        return False