Skip to main content
Guardrails are a policy mode in Limits that evaluate text (e.g. model output or user messages) for safety, PII, moderation, jailbreak detection, and URL filtering. You define a guardrails policy in the platform, then evaluate it with the SDK guard() or the API so that content is allowed, blocked, or escalated before it reaches users.

​
What guardrails cover

CategoryDescription
PIIMask or block personally identifiable information (email, phone, SSN, credit card, location, etc.) in text.
ModerationSexual content, hate/harassment, self-harm, violence, illicit activities.
JailbreakDetection of attempts to bypass system instructions.
NSFWNot-safe-for-work content filtering.
Prompt injectionDetection of prompt injection in user or model content.
URL filterAllow or block URLs by scheme and allow-list.
You create and edit guardrails in the platform via the Guardrail flow or the AI Assistant.

Policies: Guardrails

Full guide to guardrail types (PII, moderation, etc.).

SDK: guard()

Evaluate guardrails from your app.

​
How to create a guardrails policy

  1. From the Policies page (or any header with New Policy), click New Policy.
  2. In the dialog, describe your guardrails in plain language (e.g. “Mask email and phone numbers, block hate speech”).
  3. Choose Guardrail as the mode (alongside Conditions and Instructions).
  4. Submit. You are taken to /assistant/guardrail where the AI Assistant generates the guardrails configuration.
  5. When streaming finishes, you see the Results area and Simulation sidebar. Adjust which guardrails are enabled (PII, moderation, etc.), run simulation with sample text, then Deploy to save.

​
Policy structure (guardrails)

Every guardrails policy has:
ElementDescription
Policy keyUnique identifier (e.g. content-guardrail). Required.
TagsOptional labels (e.g. safety, pii).
StatusActive (enforcing) or Inactive (disabled).
You configure which guardrails are enabled (PII mask/block, moderation, jailbreak, NSFW, prompt injection, URL filter) in the Guardrail editor.

​
Editing a guardrails policy

  1. Open Policies from the sidebar.
  2. Find the policy (filter or search by key). Guardrail policies show mode Guardrail.
  3. Click Edit. You are taken to /policies/[id]/edit/guardrail.
  4. Edit which guardrails are enabled (PII types, moderation categories, etc.), run Simulation with sample text to test, then save (Deploy / Update).
You can also edit tags and status (Active/Inactive) from the policy row or the edit page.

​
Simulation

Use Simulation to test a guardrails policy before or after deploying:
  1. On the policy edit page (Guardrail tab), open the Simulation sidebar.
  2. Enter sample text (e.g. a model response or user message that might contain PII or unsafe content).
  3. Run the simulation. The result shows allow, block, or escalate and the reason.
This helps you verify that PII is masked or blocked and that moderation rules behave as expected.

​
Where guardrails appear in the platform

LocationWhat you do
PoliciesList all policies; guardrail policies have mode Guardrail. Edit, delete, or change status/tags.
New PolicyCreate a new policy; choose Guardrail mode and describe intent → you are taken to the Assistant to generate your guardrails.
Policy edit → Guardrail (/policies/[id]/edit/guardrail)Edit your guardrails settings, simulate, and deploy.

​
Summary

  • Guardrails = policy mode that evaluates text for PII, moderation, jailbreak, NSFW, prompt injection, and URL rules.
  • Create: New Policy → Guardrail + describe → Assistant generates configuration → Deploy.
  • Edit: Policies → select policy → Edit → Guardrail tab; adjust your guardrails and simulate.
  • Use in app: SDK guard(policyKeyOrTag, text) or API POST /api/policies/:key/evaluate/guardrails. See SDK Guardrails.

SDK Guardrails

Use guard() to evaluate guardrails in your app.

Policies

Create and manage all policy types.

AI Assistant

Generate policies with natural language.