Apply Guardrail

Run an LLM-backed content check that fails closed on error — block, allow, or raise

apply_guardrail runs a single LLM-backed content check and, since it fails closed by default, blocks content whenever the check itself cannot complete.

from praisonaiagents import Agent
from praisonai.capabilities import apply_guardrail

def policy_check(output: str):
    result = apply_guardrail(output, rules=["no_pii", "no_profanity"])
    return (result.passed, output if result.passed else "Blocked by guardrail.")

agent = Agent(
    name="safe-assistant",
    instructions="Draft a customer-facing reply.",
    guardrails=policy_check,
)
agent.start("Write a reply to the customer.")

The agent produces output, the guardrail LLM checks it against your rules, and the content is blocked, allowed, or the error is raised — depending on on_error.

Quick Start

1. Use it as an agent guardrail

   Wrap apply_guardrail in a validator and pass it to guardrails=.

   from praisonaiagents import Agent
   from praisonai.capabilities import apply_guardrail
   
   def validator(output: str):
       result = apply_guardrail(output, rules=["no_pii"])
       return (result.passed, output)
   
   agent = Agent(
       name="assistant",
       instructions="Answer the user politely.",
       guardrails=validator,
   )
   agent.start("Summarise the account status.")

2. Call it directly

   Run a one-off check on any string.

   from praisonai.capabilities import apply_guardrail
   
   result = apply_guardrail("Some content", rules=["no_pii", "no_profanity"])
   print(result.passed)
   print(result.violations)

3. Await the async twin

   Use aapply_guardrail inside an event loop.

   import asyncio
   from praisonai.capabilities import aapply_guardrail
   
   async def main():
       result = await aapply_guardrail("Some content", rules=["no_pii"])
       print(result.passed)
   
   asyncio.run(main())

How It Works

The check runs an LLM against your rules and returns a GuardrailResult. Any failure inside the check is routed through on_error.

Choosing on_error

on_error decides what happens when the check itself cannot produce a valid decision.

Parameters

aapply_guardrail takes the same parameters and returns the same GuardrailResult.

GuardrailResult

Both functions return a GuardrailResult.

Common Patterns

Restore the old permissive behaviour

from praisonai.capabilities import apply_guardrail

result = apply_guardrail("Some content", rules=["no_pii"], on_error="allow")

Handle the error yourself

from praisonai.capabilities import apply_guardrail

try:
    result = apply_guardrail("Some content", on_error="raise")
except Exception as exc:
    print(f"Guardrail check failed: {exc}")

Compose with an agent

from praisonaiagents import Agent
from praisonai.capabilities import apply_guardrail

def validator(output: str):
    result = apply_guardrail(output, rules=["no_pii", "no_harmful_content"])
    return (result.passed, output)

agent = Agent(name="assistant", instructions="Help the user.", guardrails=validator)
agent.start("Draft a reply.")

Best Practices

Related