Guardrails

--- title: "Enterprise Guardrails" description: "Enterprise guardrail enhancements: output-direction guardrails, enhanced dashboard management, and compliance audit integration." icon: "road-barrier" ---

Overview

Enterprise guardrails extend the OSS guardrails system with advanced capabilities for production content safety. OSS guardrails include system prompt rules, LLM-based content altering, regex blocking, PII redaction, and length limits — all available in the open-source edition. Enterprise adds output-direction guardrail support, enhanced rule configuration surfaces in the dashboard, and Enterprise compliance integration for audit trail immutability.

Guardrails intercept requests before they reach LLM providers and can optionally sanitize responses before returning to clients.

Guardrail Types

Enterprise adds enhanced guardrail capabilities beyond the OSS rule types:

Common Rule Fields

Every guardrail rule shares these optional fields:

LLM-Based Altering

The llm_based_altering guardrail rewrites selected message roles by calling an auxiliary model before the main provider request runs. This is useful for PII anonymization, prompt injection sanitization, and other content-preserving rewrites.

Configuration

guardrails:
  enabled: true
  rules:
    - name: "privacy-rewrite"
      type: "llm_based_altering"
      user_path: "/team/privacy"
      order: 1
      llm_based_altering:
        model: "gpt-4o-mini"
        roles: ["user"]
        max_tokens: 4096
        skip_content_prefix: "### safe"

Settings

How It Works

When llm_based_altering runs, Aurora calls the auxiliary model through the normal translated request path:

• Path: /v1/chat/completions
• User path: {guardrail.user_path or caller user path}/guardrails/{guardrail name}
• Request origin: guardrail

Guardrails are explicitly skipped for the internal request to avoid recursion.

Regex Block

Block or sanitize messages that match configured regular expressions.

guardrails:
  enabled: true
  rules:
    - name: "block-system-escape"
      type: "regex_block"
      order: 0
      regex_block:
        patterns:
          - "ignore all previous instructions"
          - "you are now.*"
        action: "block"

Settings

PII Redaction

Deterministically redact personally identifiable information from messages.

guardrails:
  enabled: true
  rules:
    - name: "pii-filter"
      type: "pii_redact"
      order: 0
      pii_redact:
        kinds:
          - email
          - phone
          - ssn
          - credit_card

Settings

Length Limit

Enforce maximum character or estimated token counts on messages.

guardrails:
  enabled: true
  rules:
    - name: "max-length"
      type: "length_limit"
      order: 0
      length_limit:
        max_chars: 100000
        max_estimated_tokens: 32000

Settings

Output Guardrails

Guardrails can also apply to provider responses. Configure the direction to control when rules run:

input   — Run before the request reaches the provider (default)
output  — Run on the provider response before returning to the client
both    — Run in both directions

Output guardrails use the same rule types and configuration as input guardrails.

Workflow Integration

Guardrails are enabled per-request through workflow policies. Each workflow payload controls which guardrails apply to matching requests:

curl -X POST http://your-aurora-host/admin/api/v1/workflows \
  -H "Authorization: Bearer $AURORA_MASTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "scope_provider_name": "openai_primary",
    "scope_user_path": "/team/alpha",
    "workflow_payload": {
      "schema_version": 1,
      "features": {
        "guardrails": true
      },
      "guardrails": ["privacy-rewrite"]
    }
  }'

Admin API

For endpoint reference see the Admin API section.

curl -X PUT "http://your-aurora-host/admin/api/v1/guardrails/privacy-rewrite" \
  -H "Authorization: Bearer $AURORA_MASTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "privacy-rewrite",
    "type": "llm_based_altering",
    "order": 1,
    "llm_based_altering": {
      "model": "gpt-4o-mini",
      "roles": ["user"]
    }
  }'

Reference

See OSS Guardrails for the complete guardrails pipeline reference, execution order, and system prompt modes.