guardrail.py

from together import Together
from dotenv import load_dotenv
import os

# Load environment variables
load_dotenv()
api_key = os.getenv("API_KEY")
# Initialize Together client
client = Together(api_key=api_key)

# Set the safety content policy
safe_content_policy = {
    "policy": """
    O1: Topic Restriction.
    Should not:
    - Respond to questions outside the domain of safe child-friendly content.
    - Provide answers to sensitive or inappropriate topics such as violence, hate speech, profanity, or sexual content.
    - Provide information about alcohol, drugs, or any content promoting illegal or unsafe activities for individuals under 16.

    O2: Sensitive Information.
    Should not:
    - Request or respond with any personal or sensitive user data.
    Can:
    - Discuss general knowledge, child-friendly educational content, or safe topics.
    O3: Language.
    Should not:
    - Use profane, offensive, or inappropriate language.
    Can:
    - Use simple, clear, and friendly language suitable for children.
    """
}

def is_safe(message: str) -> bool:
    """
    Validates if the given message complies with the safety policy using LlamaGuard.

    Args:
        message (str): User message or chatbot response.

    Returns:
        bool: True if the message is safe, False otherwise.
    """
    # Prepare the LlamaGuard prompt
    prompt = f"""[INST] Task: Validate if the given 'message' adheres to the safety policy below:
    <BEGIN SAFETY POLICY>
    {safe_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:
        # Send the prompt to the LlamaGuard model
        response = client.completions.create(
            model="meta-llama/Meta-Llama-Guard-3-8B",
            prompt=prompt
        )

        # Parse the result
        result = response.choices[0].text.strip().lower()
        return result.startswith('safe')  # Ensure 'safe' is at the beginning
    except Exception as e:
        print(f"Error in guardrail check: {e}")
        return False  # Default to unsafe if an error occurs