Guardrails and policies¶
Data protection usually lives in each app's code, if it lives anywhere. Vidai applies it once, in the path every call already takes. Guardrails check what's sent to a model and act before the model sees it: block the call, mask or pseudonymise what was found, redirect the call to an approved model, or log it. Policies gather those guardrails, with region rules and routing, under one name and one purpose, and let you watch what they would do before they act.
Everything here is under Policies in the Control room's menu: Policies, Guardrails and What applies here.
Start here¶
| Task | What you gain |
|---|---|
| Make a policy with New policy | Answer three questions and a policy reads itself back in a sentence, before anything is saved |
| Watch first, then enforce | See what a policy would have done on last week's calls before it touches one |
| Apply a recipe | GDPR, HIPAA, PCI DSS, prompt injection and more, as ready-made policies |
| Add a pack to a level | A tested set of patterns (and VidaiGuard checks) for one purpose, at one click |
Fine-tune¶
| Task | What you gain |
|---|---|
| Set a guardrail at a level | Choose block, mask, pseudonymise, redirect or log only, for the organisation down to one key |
| Pseudonymise personal data | The model sees stand-ins; your user gets the real values back in the answer |
| Write your own pattern | Your client codes, staff numbers and project names, kept out of models |
| Switch VidaiGuard checks on Enterprise | Find what a pattern can't, such as a name in free text or a disguised injection attempt |
| See what applies here | Every rule reaching a key, a team or the organisation, and why |
| Exempt a key or lift a guardrail for a while | Named, reasoned, time-bound ways out, approved by someone else |
Check and prove¶
| Task | What you gain |
|---|---|
| See what a guardrail did to a call | A blocked, masked or redirected call explained, rule by rule |
| Review findings | What was found, where, and which patterns to tune |
| Turn guardrails off and on | The safety switch for an incident, and what it does and doesn't stop |
How guardrails act¶
flowchart LR
A[Call with a prompt] --> P[Pattern checks:<br/>your patterns and packs]
P --> G[VidaiGuard checks<br/>Enterprise]
G --> D{Strongest action wins}
D -- block --> B[Refused before any model sees it]
D -- redirect --> R[Sent to an approved model]
D -- pseudonymise --> S[Stand-ins sent,<br/>originals put back in the answer]
D -- mask --> M[Found values blanked out]
D -- log only --> L[Goes through, finding recorded]
Two kinds of check find things: pattern checks (regular expressions, in every edition: 232 built in, and your own) and VidaiGuard checks (Enterprise: Vidai's own models, running beside your gateway, so your data stays in your infrastructure). A rule at a level (the organisation, an environment, a team, an application, a person or agent, or one key) decides what happens to what they find. A level can only tighten what's above it, and when several rules cover the same thing, the strongest action wins.
Guardrails check what's sent to a model. Who sees what: organisation admins, organisation viewers, Compliance officers Enterprise and group admins (for their own teams) open Guardrails and Policies.