Guardrail Nodes

Overview

Guardrail nodes check input or output for safety concerns. They have two output handles: pass and block.

{
  "id": "guard-1",
  "type": "guardrail",
  "data": {
    "type": "guardrail",
    "guardrailType": "content",
    "config": {
      "blockedTopics": ["violence", "illegal"],
      "action": "block"
    }
  }
}

PII Detection

Detects personally identifiable information (emails, phone numbers, SSNs, etc.).

{
  "guardrailType": "pii",
  "config": {
    "detectors": ["email", "phone", "ssn", "credit_card", "address"],
    "action": "redact",
    "redactWith": "[REDACTED]"
  }
}

Action	Behavior
block	Reject the message entirely
redact	Replace PII with placeholder text
warn	Allow but flag for review

Jailbreak Prevention

Detects prompt injection and jailbreak attempts.

{
  "guardrailType": "jailbreak",
  "config": {
    "sensitivity": "medium",
    "action": "block",
    "customPatterns": []
  }
}

Sensitivity levels: low, medium, high

Content Filtering

Filters content based on blocked topics or custom rules.

{
  "guardrailType": "content",
  "config": {
    "blockedTopics": ["violence", "illegal", "adult"],
    "action": "block",
    "customRules": [
      { "pattern": "competitor_name", "action": "warn" }
    ]
  }
}

Placement

Place guardrails before agent nodes to check input, or after to check output:

Start → Guardrail (input check) → Agent → Guardrail (output check) → End