Guardrails and Safety

LLM Safety and Guardrails

# Input validation
def validate_input(text: str) -> tuple[bool, str]:
    if len(text) > 2000:
        return False, 'Input too long (max 2000 chars)'
    injection_patterns = [
        'ignore previous', 'new instructions',
        'system prompt', 'jailbreak',
    ]
    for p in injection_patterns:
        if p.lower() in text.lower():
            return False, 'Invalid input detected'
    return True, ''

# Output validation
def validate_output(response: str, allowed_topics: list) -> str:
    # Check relevance to allowed topics
    for topic in allowed_topics:
        if topic.lower() in response.lower():
            return response
    return 'I can only help with questions about our products.'

# Guardrails AI library
from guardrails import Guard
from guardrails.hub import ToxicLanguage, DetectPII
guard = Guard().use(ToxicLanguage).use(DetectPII)
result = guard.validate(llm_output)

# Content moderation API
# OpenAI moderation endpoint (free)
response = client.moderations.create(input=user_message)
if response.results[0].flagged:
    return 'Content policy violation'