AI guardrails on Rafay — for customers
A short guide for the customer turning on Votal Shield guardrails for their AI workloads on Rafay. No code required.
What guardrails do for you
When your people and apps use AI (chat, assistants, agents), guardrails inspect each prompt and each model response and act on your policy. They can:
- Stop sensitive data leaving — Social Security and card numbers, secrets, API keys, and other protected data in a prompt.
- Block prompt injection and unsafe content — manipulation attempts and outputs your policy disallows.
- Govern agents and tools — what an AI agent is allowed to do and see.
Your policy is yours alone. It is applied only to your organization’s traffic and is isolated from every other customer.
Turning it on
In the Rafay console, enable Votal Shield guardrails for your AI workloads. Rafay provisions your dedicated guardrail tenant automatically — there is nothing to install.
Choosing your policy
You decide which guardrails are active. Depending on how Rafay surfaces it, you configure this either in the Shield policy portal or directly in the Rafay UI. Options include:
- Built-in guardrails — PII and secret detection, toxicity, prompt-injection defense, and more.
- Custom policies — describe a rule in plain language (for example, “block messages containing customer account numbers”) and choose whether to warn, redact, or block.
- Framework-aligned policies — turn on sets mapped to common AI risk frameworks.
Changes take effect on the next request. No redeploy.
Monitor first, then enforce
- Monitor — guardrails observe and record decisions without blocking. Use this to see what your traffic looks like.
- Enforce — guardrails block policy violations in real time.
Start in monitor, review what would have been blocked, then switch to enforce.
What a user sees when something is blocked
In enforce mode, a prompt that violates your policy is stopped before it reaches the AI model — the sensitive content never leaves. The person sees that their message was blocked by policy.
Seeing what happened
Every decision — allowed, blocked, or redacted — is recorded with the policy that fired, visible in your telemetry/dashboard. Use it to tune policy and to show auditors what was enforced.
Where it runs and your data
- Hosted — Rafay calls the Votal Shield service.
- In your cluster — for strict data residency, Shield (and optionally its model) runs inside your own environment, so prompts are inspected without leaving it.
Either way, the guardrail decision is based only on the content being screened and your policy; your model credentials are never shared with Shield.
Getting help
Your Rafay administrator can adjust which guardrails are on, switch between monitor and enforce, and pull decision reports. For policy design questions, Rafay can bring in Votal.