AI Agent1 input → 2 outputs

Guardrails

Applies intelligent safety guardrails to AI responses using Regex patterns or AI-enhanced logic for PHI redaction, content moderation, and quality checks.

Inputs

payload string | object: The content to process. If Input Property is configured, that property is checked instead.
aiagent object: AI configuration (from AI Model node). Required for AI-Enhanced mode.
toolInput string | object: Automatic fallback input source if connected after an AI Tool Adapter logic.
agent_error object: If present (from upstream AI error), the node automatically routes to Fail (Output 2).

Outputs

Pass (Safe/Redacted)

payload string | object

The safe content. If input was JSON, the structure is preserved (with redactions applied).

guardrails object

Metadata: { mode: 'ai', verdict: 'pass', redactedCount: 2 }
Fail (Block/Violation)

payload.error string

Failure message (e.g., "GUARDRAIL BLOCK: Policy Violation"). Suitable for returning to AI.

guardrails object

Violation details: { verdict: 'fail', reason: '...' }

Configuration

Mode

Regex-Only (Fast): Local pattern matching (PHI, Custom Regex). No API calls.
AI-Enhanced (Intelligent): Uses LLM to reason about context, tone, and moderation.

Rules (AI Mode)

PHI Redaction: Intelligent de-identification.
Content Moderation: Block harmful/illegal content.
Tone Check: Rewrite for professional tone.
Custom Rule: User-defined logic with templating (e.g. {{msg.checkType}}).

Typical Usage

Healthcare Agent Flow

┌────────────┐   ┌──────────┐   ┌────────────────┐
│ User Query │──▶│ AI Model │──▶│    AI Agent    │
└────────────┘   └──────────┘   └───────┬────────┘
                                        │
           ┌────────────────────────────┘
           ▼
┌──────────────┐   ┌─────────────────┐   ┌────────────┐
│  Guardrails  │──▶│ Update History  │──▶│  Response  │
└──────┬───────┘   └─────────────────┘   └────────────┘
       │
       ▼ (Output 2: Block/Fail)
  ┌──────────┐
  │  BLOCKED │
  └──────────┘

Examples

1. Redaction (Regex or AI)

Input:  "Patient John Doe (DOB 01/01/1980) has diabetes."
Output: "Patient [REDACTED] (DOB [REDACTED]) has diabetes."

2. Blocking (Content Moderation)

Input:  "I hate you and want to hurt someone."
Output (Port 2):
{
  "payload": {
    "error": "GUARDRAIL BLOCK: Hate speech detected. Please rewrite..."
  }
}

3. Custom Rule with Templating

Rule: "Block if text mentions {{msg.bannedFruit}}"

msg.bannedFruit = "apple";
Input: "I like apple pie."
Output (Port 2): "GUARDRAIL BLOCK: Text mentions apple."

Best Practices

Healthcare: Always place Guardrails after AI Agent.
Performance: Use Regex-Only for high-volume data.
Fail-Safe: Connect Output 2 to an error handler or back to the AI (with a prompt to retry).
Testing: Verify custom rules with known bad inputs before production.

┌────────────┐ ┌──────────┐ ┌────────────────┐ │ User Query │──▶│ AI Model │──▶│ AI Agent │ └────────────┘ └──────────┘ └───────┬────────┘ │ ┌────────────────────────────┘ ▼ ┌──────────────┐ ┌─────────────────┐ ┌────────────┐ │ Guardrails │──▶│ Update History │──▶│ Response │ └──────┬───────┘ └─────────────────┘ └────────────┘ │ ▼ (Output 2: Block/Fail) ┌──────────┐ │ BLOCKED │ └──────────┘