Guardrails Tools

The Guardrails provider validates content flowing through a workflow before it is stored, returned, or passed downstream. A single guardrails_validate tool can check whether content is valid JSON, matches a regex pattern, is grounded in a knowledge base (hallucination check via RAG + LLM scoring), or contains personally identifiable information (PII) that should be blocked or masked. Use it as a safety gate between a generation step and an output/storage step.

Overview

Operations

The provider exposes one tool whose behavior branches on validationType. The four modes are:

• json — passes if the input parses as valid JSON.
• regex — passes if the input matches the supplied regex pattern.
• hallucination — queries a knowledge base (RAG, topK chunks), then uses an LLM to score how well the input is supported on a 0–10 confidence scale; passes if the score is at or above threshold.
• pii — detects PII entities; in block mode it fails when PII is found, in mask mode it passes and returns the text with each entity replaced by a <TYPE> placeholder.

Configuration

guardrails_validate

Supported piiEntityTypes ids: PERSON, EMAIL_ADDRESS, PHONE_NUMBER, INDIAN_MOBILE_NUMBER, LOCATION, DATE_TIME, IP_ADDRESS, URL, CREDIT_CARD, IBAN_CODE, CRYPTO, MEDICAL_LICENSE, NRP, IN_AADHAAR, IN_PAN, IN_VEHICLE_REGISTRATION, IN_VOTER, IN_PASSPORT, IN_BANK_ACCOUNT, IN_DRIVER_LICENSE, IN_GST.

Outputs

guardrails_validate

• passed (boolean) — Whether validation passed.
• validationType (string) — The type of validation that was performed.
• input (string) — The original input that was validated.
• score (number, optional) — Confidence score from 0–10. Present only for hallucination checks.
• reasoning (string, optional) — Explanation of the confidence score. Present only for hallucination checks.
• detectedEntities (array, optional) — Detected PII entities, each with type, start, end, score, and text. Present only for pii detection.
• maskedText (string, optional) — Text with PII replaced by <TYPE> placeholders. Present only for pii detection in mask mode.
• error (string, optional) — Error message when validation fails.

YAML Example

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

A hallucination-check example:

guardrails_2:
  type: guardrails
  name: "Guardrails"
  inputs:
    operation: "guardrails_validate"
    input: "{{agent_1.content}}"
    validationType: "hallucination"
    knowledgeBaseId: "kb_abc123"
    model: "gpt-4o-mini"
    threshold: 3
    topK: 5
    apiKey: "{{OPENAI_API_KEY}}"
  connections:
    outgoing:
      - target: next-block-id

Tips

• The validation behavior is selected entirely by validationType. Supply only the parameters that mode needs: regex for regex checks, knowledgeBaseId / model / apiKey for hallucination checks, and piiMode / piiEntityTypes / customPiiPatterns for PII checks.
• For pii masking, leave piiEntityTypes empty to scan for every built-in entity type, then read {{guardrails_1.maskedText}} downstream — it replaces each detection with a <TYPE> placeholder. In block mode the block fails (with error listing the detected entity counts) instead of returning masked text.
• The hallucination check only needs an apiKey when the selected model is a non-hosted provider model (or a non-Ollama model when self-hosting). Reference the key with an environment variable, e.g. apiKey: "{{OPENAI_API_KEY}}", and never hardcode it. For Azure OpenAI models, also set azureEndpoint and azureApiVersion.
• A low hallucination score means the input was not well supported by the knowledge base; if no relevant chunks are retrieved, scoring falls back to general knowledge and stays deliberately strict, so make sure the knowledge base actually contains the domain content you are validating against.