Skip to content

Guardrails: what a call may contain

The problem. Data minimisation is easy to write and hard to enforce: every app would need its own filters, kept up to date, and nobody can show they ran. With Vidai, guardrails check what's sent to a model in the gateway, on every call, and act before the model sees it. Each finding is recorded against the call and the policy that decided it.

What they find

  • Pattern checks: regular expressions in categories, such as personal data, card numbers (checked with their check digits), health identifiers, secrets and keys, prompt injection, profanity and your own internal terms. Ready-made packs group them by purpose: GDPR, HIPAA, PCI DSS, credentials, agent safety, and personal data packs for regions (the UK, India, the UAE, Saudi Arabia, Qatar, Bahrain, Oman, Kuwait and the GCC among them). You can write your own patterns and test them on sample text first.
  • VidaiGuard checks Enterprise: once VidaiGuard is set up, Vidai's own models, running beside your gateway, also find what a pattern can't, such as a name in free text or a disguised injection attempt. Your data stays in your infrastructure.

Guardrails are complete on pattern checks alone, with every action below: on Community, and on Enterprise while VidaiGuard isn't set up. Packs and policies work the same either way; their VidaiGuard checks start once it's set up.

What they do

Five actions, strongest first. When several rules cover the same thing, the strongest wins, and a lower level (a team, an application, a key) can only tighten what's set above it.

Action What happens to the call
Block The call is refused, and the app is told why
Redirect The call goes to another approved model instead
Pseudonymise Found values are swapped for stand-ins; your user gets the real values back in the answer
Mask Found values are blanked out before the model sees them
Log only The call goes through; the finding is recorded

Log only stores the original

A guardrail that only logs personal data (cards, health data, ID numbers) says Stores the original text: wherever call bodies are kept, the original is stored. Mask or block to keep it out.

Observe before you enforce

Every guardrail can start by watching: it records what it would have done and changes nothing. A named policy that's observing shows What enforcing would have done over the last seven days, so you switch to enforcing knowing what will change. See Named compliance policies.

Findings to review

Policies › Guardrails › Findings lists what was found, by category and pattern. Mark false positives, and see which patterns need tuning. Each finding also shows on its call in Request logs, with the policy that decided it.

Good to know

  • Guardrails check requests, what's sent to a model; answers coming back aren't checked.
  • Where VidaiGuard is set up but isn't answering, degrade (the default) lets calls through on pattern checks alone (and marks them partly checked); strict refuses calls that needed a VidaiGuard check. Strict isn't offered while VidaiGuard isn't set up. Each pack says what runs for it in one badge, such as Patterns only, VidaiGuard not set up or VidaiGuard unavailable: patterns only: see Add a pack to a level.