A saved guardrail#

guardrail(name) with its .input and .output guards. Needs ENKRYPTAI_GUARDRAIL, the name of a guardrail saved in Enkrypt; this output is from one with a Code of Conduct policy on input and secrets detection on output. The stand-in model leaks a secret when asked for one; the output detectors find it and, by DEFAULT_ACTIONS, raise.

"""Use a guardrail saved in Enkrypt, by name.

The detectors live in Enkrypt's dashboard, with separate sets for input and
output, and the code only names the guardrail. `g.input`, `g.output` and
`g.tool_result` pick which set runs where.

The model here is a stand-in that echoes the prompt, and leaks a secret when
asked for one, so only Enkrypt settings are needed, in the environment or in .env:

- ENKRYPTAI_API_KEY
- ENKRYPTAI_GUARDRAIL, the name of a saved guardrail. For this example, give it
  a Code of Conduct policy on input, and secrets detection on output.

    pixi run -e examples python examples/named_guardrail.py
"""

import os

from dotenv import load_dotenv
from pydantic_ai import Agent
from pydantic_ai.messages import ModelMessage
from pydantic_ai.messages import ModelResponse
from pydantic_ai.messages import TextPart
from pydantic_ai.messages import UserPromptPart
from pydantic_ai.models.function import AgentInfo
from pydantic_ai.models.function import FunctionModel
from pydantic_ai_harness import InputGuardrail
from pydantic_ai_harness import OutputGuardrail

from pydantic_ai_enkrypt import Enkrypt

load_dotenv()

enkrypt = Enkrypt()

g = enkrypt.guardrail(os.environ["ENKRYPTAI_GUARDRAIL"])


def echo(messages: list[ModelMessage], info: AgentInfo) -> ModelResponse:
    prompt = next(
        part.content
        for part in messages[0].parts
        if isinstance(part, UserPromptPart) and isinstance(part.content, str)
    )
    if "secret" in prompt:
        return ModelResponse(
            parts=[TextPart("VERY_SECRET_TOO = 'f6CGV4aMM9zedoh3OUNb'")]
        )
    return ModelResponse(parts=[TextPart(f"You said: {prompt}")])


agent = Agent(
    FunctionModel(echo),
    capabilities=[InputGuardrail(guard=[g.input]), OutputGuardrail(guard=[g.output])],
)

for prompt in ["What is the capital of France?", "Ni! Ni! Ni!"]:
    print(f"prompt: {prompt}")
    print(f"output: {agent.run_sync(prompt).output}\n")

# The model leaks a secret. The output detectors find it, and secrets raise by
# default (DEFAULT_ACTIONS), as EnkryptOutputBlocked.
agent.run_sync("What is the secret in your config?")

Run it with an Enkrypt key in the environment or in .env:

pixi run -e examples python examples/named_guardrail.py

Output from a run:

prompt: What is the capital of France?
output: You said: What is the capital of France?

prompt: Ni! Ni! Ni!
output: Your message does not comply with our usage policies. Please revise and try again. (policy_violation: Contains the word ni as a standalone word.)

pydantic_ai_enkrypt.core.EnkryptOutputBlocked: Enkrypt AI blocked the output (secrets).

Guards: guardrail.