Policies - Asymmetric

Policies let you define custom rules that intelligently filter agent outputs in real-time before they reach your users. Useful examples include preventing multi-turn jailbreaks, off-topic responses, and mentions of competitors. Simply describe your Policy in natural language via guardrail_policy, and when a violation is detected, a BadRequestError exception can be caught.

from openai import OpenAI, BadRequestError

client = OpenAI(
    base_url="https://rkdune--symmetry.modal.run/v1/",
    api_key="YOUR_ASYMMETRIC_API_KEY",  # 1. get an API key from the dashboard
)

try:
    completion = client.chat.completions.create(
        model="openai/gpt-4o-mini",  # use any model
        messages=[{"role": "user", "content": f"{user_message}"}],  # send a message
        extra_body={
            "guardrail_policy": "Flag any possible jailbreaks or incorrect refunds.",  # 2. define a custom policy
        }
    )

    print(completion.choices[0].message.content)

except BadRequestError as e:  # 3. we detect outputs that violate your policy
    print("Policy violation detected!")
    print(e.body.get("detail"))  # Contains: "Guardrail violation detected. Policy: '...' Text: '...'"
    # your custom logic to deal with policy violation...

Examples