Guardrails

The Guardrails block checks content against one validation type and reports whether it passed. Use it to catch malformed JSON, off-pattern text, ungrounded answers, or PII before the content moves on. Each block runs one check; chain several to apply more than one.

Validation Types

Valid JSON

Checks that the content parses as valid JSON. Use it before a Function or downstream block reads a model's structured output.

• <guardrails.passed> — true if the content is valid JSON
• <guardrails.error> — the parse error when it isn't, like Invalid JSON: Unexpected token

Regex Match

Checks the content against a regular expression — an email, a phone number, a URL, or any pattern you define.

• Regex Pattern — the expression to match, like ^[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}$ for an email
• <guardrails.passed> — true if the content matches; <guardrails.error> otherwise

Hallucination Check

Scores how well AI output is grounded in a knowledge base. The block retrieves relevant context, sends it to a model with the output, and the model returns a confidence score from 0 (completely ungrounded) to 10 (fully supported). Validation passes when the score meets the threshold.

• Knowledge Base — the knowledge base to check against
• Model — the model that scores grounding. Use a strong reasoning model; the default is claude-sonnet-4-6. The API key is supplied for you on hosted Sim
• Confidence — the minimum score to pass, from 0 to 10 (default 3)
• Top K (advanced) — how many knowledge base chunks to retrieve (default 5)

Outputs <guardrails.score> (0–10) and <guardrails.reasoning> (why the model scored it that way) alongside passed.

PII Detection

Detects personally identifiable information with Microsoft Presidio — over 30 entity types across several countries and languages.

• PII Types to Detect — pick the entity types from the modal, grouped by region:
  • Common — person name, email, phone, credit card, IP address, and more
  • USA — SSN, driver's license, passport, bank account, ITIN
  • UK — NHS number, national insurance number
  • Spain — NIF, NIE · Italy — fiscal code, driver's license, identity card, passport · Poland — PESEL · Singapore — NRIC/FIN
  • Australia — ABN, ACN, TFN, Medicare · India — Aadhaar, PAN, vehicle registration, voter number, passport
• Action — Block fails validation when any selected type is found (default); Mask also replaces the PII with masked values
• Language — the detection language (default English)

Outputs <guardrails.detectedEntities> (each with type, location, and confidence) and, in Mask mode, <guardrails.maskedText>. passed is false when any selected PII is found.

Configuration

Content to Validate

The input to check. Usually an earlier output like <agent.content>, <function.result>, or an API response.

Validation Type

Which of the four checks to run: Valid JSON, Regex Match, Hallucination Check, or PII Detection.

Outputs

Every validation type returns:

Hallucination adds <guardrails.score> and <guardrails.reasoning>; PII adds <guardrails.detectedEntities> and <guardrails.maskedText>.

Examples

Validate JSON before parsing

Check the Agent's output is valid JSON, then branch on <guardrails.passed> before a Function parses it.

Prevent hallucinations

Score the answer against a knowledge base, then gate on <guardrails.score> to send a grounded answer or flag a weak one.

Block PII in user input

Detect PII in the input and branch on <guardrails.passed> to process clean input or reject it.

Best Practices

• Branch on the result. Read <guardrails.passed> in a Condition to route valid and invalid content down different paths.
• Validate JSON before you parse it. A check upstream is cheaper than a parse error in a Function block.
• Pick only the PII types you need. Selecting fewer entity types keeps detection fast and focused.
• Tune the hallucination threshold. Raise the confidence floor for stricter grounding, lower it to allow more latitude.
• Mask when you log. Use Mask mode for content you store or log, so PII never lands in plain text.
• Chain checks. One block runs one type, so place several in sequence to validate format and then scan for PII.