AAPP-871 Guardrail

[bigdata2]

Add Guardrails client with tracing integration

Adds a Guardrails client to the ADK that allows users to evaluate content against safety rails directly from their agents.

What's included

• gradient_adk/guardrails.py — Async Guardrails client with a single check() method for evaluating content against jailbreak, content moderation, and sensitive data rails
• Tracing integration — Guardrail evaluations are automatically captured as tool spans in the ADK trace when used inside @entrypoint
• __init__.py — Exports Guardrails as a public API
• Unit tests — 12 tests covering check(), error handling, and tracing
• README — Updated with guardrails documentation, usage examples, and rail type reference

Usage

from gradient_adk import entrypoint, RequestContext, Guardrails

guardrails = Guardrails()

@entrypoint
async def main(input: dict, context: RequestContext):
    result = await guardrails.check(
        rail_type="jailbreak",
        messages=[{"role": "user", "content": input["prompt"]}],
    )
    if not result["allowed"]:
        return {"error": "Blocked", "violations": result["violations"]}
    return await process(input["prompt"])

Available rail types

Rail Type	Description
jailbreak	Detects prompt injection and jailbreak attempts
content_moderation	Detects harmful, violent, or inappropriate content
sensitive_data	Detects PII and sensitive information

Testing

• Unit tests: 12 tests, all passing
• Local E2E: Tested via gradient agent run against preview guardrails service
• Preview E2E: Jailbreak and content moderation rails verified (safe + unsafe prompts)
• Production E2E: Deployed and verified at guardrails.do-ai.run/v2/rail