Guardrails Block

The Guardrails block validates content against a chosen rule before it moves downstream. It can check whether content is valid JSON, matches a regex pattern, detect hallucinations using RAG + LLM scoring, or detect (and optionally mask) PII. Reach for it when you need to gate or sanitize text inside a workflow.

Overview

When to Use

• Verify that an upstream block produced syntactically valid JSON before parsing it downstream.
• Confirm content matches a specific format (email, phone, URL, date, custom pattern) via a regex rule.
• Fact-check an AI response against a knowledge base and score it for hallucinations using an LLM.
• Detect personally identifiable information (names, emails, phone numbers, Aadhaar, PAN, credit cards, etc.) for privacy compliance.
• Mask PII in place so sensitive data is scrubbed before storage or sharing.
• Block requests that contain disallowed personal information before they reach downstream blocks.

Configuration

Content to Validate

The input field holds the text to validate. Accepts {{blockName.field}} references to upstream block outputs. Required for all validation types.

Validation Type

The validationType dropdown selects the guardrail rule to apply. Required. Drives which additional fields are shown.

Regex Pattern

Shown when validationType is regex. The regex field holds the regular expression to match against (e.g. ^\{.*\}$ for any JSON object, or ^[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}$ for email). Required when this validation type is selected.

Knowledge Base

Shown when validationType is hallucination. The knowledgeBaseId selector picks a knowledge base that supplies factual context for the hallucination check. Required when this validation type is selected.

AI Model

Shown when validationType is hallucination. The model combobox selects the LLM used for confidence scoring (default gpt-4o-mini). Required when this validation type is selected.

Confidence Threshold

Shown when validationType is hallucination. The threshold slider (0–10, default 3) sets the minimum score a response must achieve to pass. Responses scoring below this value are marked as failing.

Number of Context Chunks

Shown when validationType is hallucination (advanced mode). The topK slider (1–20, default 5) controls how many knowledge base chunks are retrieved to build the fact-checking context.

API Key

Shown when validationType is hallucination and the selected model requires a provider key. The apiKey field supplies the API key for the LLM provider. Use {{MY_API_KEY}} to reference a workspace secret. Required unless a hosted or local model is selected.

Azure OpenAI Endpoint

Shown when validationType is hallucination and an Azure OpenAI model is selected. The azureEndpoint field holds the endpoint URL (e.g. https://your-resource.openai.azure.com).

Azure API Version

Shown when validationType is hallucination and an Azure OpenAI model is selected. The azureApiVersion field holds the API version string (e.g. 2024-07-01-preview).

PII Types to Detect

Shown when validationType is pii. The piiEntityTypes checkbox-list selects which entity types to detect. Leave all unchecked to detect every supported type. Available options:

Custom PII Patterns

Shown when validationType is pii (advanced mode). The customPiiPatterns field accepts a JSON object mapping custom type names to regex patterns, e.g. {"EMPLOYEE_ID": "EMP\\d{6}"}. Do not use ^ or $ anchors — patterns match within text, not the entire string.

PII Action

Shown when validationType is pii. The piiMode dropdown sets what happens when PII is found. Required.

Language

Shown when validationType is pii. The piiLanguage dropdown sets the detection language (default en).

Inputs & Outputs

• Inputs:

  • input (string) — Content to validate
  • validationType (string) — Type of validation to perform (json, regex, hallucination, or pii)
  • regex (string) — Regex pattern for validation (used when validationType is regex)
  • knowledgeBaseId (string) — Knowledge base ID for hallucination check
  • threshold (number) — Confidence threshold for hallucination check (0–10)
  • topK (number) — Number of context chunks to retrieve from the knowledge base
  • model (string) — LLM model for hallucination scoring
  • apiKey (string) — API key for the model provider
  • azureEndpoint (string) — Azure OpenAI endpoint URL
  • azureApiVersion (string) — Azure API version
  • piiEntityTypes (json) — PII entity types to detect
  • piiMode (string) — PII action mode (block or mask)
  • piiLanguage (string) — Language for PII detection
  • customPiiPatterns (json) — Custom PII patterns as a JSON object

• Outputs:

  • passed (boolean) — Whether validation passed
  • validationType (string) — Type of validation performed
  • input (string) — Original input that was validated
  • score (number) — Confidence score (0–10; only for hallucination check)
  • reasoning (string) — Reasoning for the confidence score (only for hallucination check)
  • detectedEntities (json) — Detected PII entities, each with type, start, end, score, and text (only for PII detection)
  • maskedText (string) — Text with PII replaced by type placeholders (only for PII detection in mask mode)
  • error (string) — Error message if validation failed

Tools

Guardrails Validate (guardrails_validate) — Validates content using one of four strategies: JSON syntax check, regex pattern matching, hallucination scoring via RAG + LLM, or PII detection with optional masking. Calls the internal /api/tools/guardrails endpoint and returns a structured result with passed, optional score/reasoning for hallucination checks, and detectedEntities/maskedText for PII checks.

YAML Example

guardrails_1:
  type: guardrails
  name: "Guardrails"
  inputs:
    input: "{{agent_1.output.content}}"
    validationType: "pii"
    piiEntityTypes:
      - "PERSON"
      - "EMAIL_ADDRESS"
      - "PHONE_NUMBER"
    piiMode: "mask"
    piiLanguage: "en"
  connections:
    outgoing:
      - target: next-block-id