Guardrails AI — Input & Output Safety¶

from guardrails import Guard
from guardrails.errors import ValidationError
from guardrails_ai.secrets_present import SecretsPresent
from guardrails_ai.valid_length import ValidLength
from guardrails_ai.valid_choices import ValidChoices

guard = (
    Guard(name="support-bot")
    .use(SecretsPresent(on_fail="exception"), ValidLength(max=4000, on_fail="exception"), on="messages")
    .use(ValidChoices(choices=["bug", "feature", "question"], on_fail="reask"))           # on="output"
)


def answer(user_text: str) -> str:
    try:
        outcome = guard(model="anthropic/claude-haiku-4-5",
                        messages=[{"role": "user", "content": user_text}])
    except ValidationError:
        return "Please remove credentials from your message."     # the LLM was never called
    return outcome.validated_output


answer("my key is AKIAIOSFODNN7EXAMPLE, why 500?")                  # blocked by the input guard

from guardrails_ai.detect_pii import DetectPII
from guardrails_ai.toxic_language import ToxicLanguage
from guardrails_ai.competitor_check import CompetitorCheck

output_guard = Guard(name="support-answer").use(
    DetectPII(pii_entities=["EMAIL_ADDRESS", "PHONE_NUMBER", "CREDIT_CARD"], on_fail="fix"),
    ToxicLanguage(threshold=0.5, validation_method="sentence", on_fail="fix"),
    CompetitorCheck(competitors=["Acme Cloud", "Globex"], on_fail="fix"),
)

from guardrails_ai.provenance_llm import ProvenanceLLM

rag_guard = Guard(name="docs-rag").use(
    ProvenanceLLM(validation_method="sentence", llm_callable="openai/gpt-4o-mini", on_fail="fix"),
)
outcome = rag_guard.validate(answer, metadata={"sources": [chunk.text for chunk in retrieved_chunks]})