PII Redactor — Mask Before the LLM (Reversible)

Strip emails, phone numbers, credit cards and other personal data out of your text before it reaches an AI model — then restore it afterwards with a reversible token map. GDPR-friendly redaction for AI, automation and analytics pipelines.

What it does

• Detects and masks PII — email addresses, phone numbers, credit card numbers, US SSNs, IP addresses, IBANs and URLs.
• Reversible tokens — each value becomes a typed token like [EMAIL_1] or [CREDITCARD_1], so the text stays readable and the structure survives.
• One combined token map — a single reversible map (token → original value) is written to the run's key-value store, so a later step can put the real values back.
• Two modes — mask (sequential typed tokens) or hash (typed tokens with a stable hash).
• Privacy by default — the original text is never stored or logged unless you explicitly opt in.

Use cases

• Mask before the LLM — redact prompts before an AI node (OpenAI, Anthropic, etc.), then restore names, emails and numbers in the model's output. Keep raw PII out of third-party models.
• n8n / Make / Zapier flows — drop it in as a step: redact before the AI node, restore after. Great for support-ticket, email and document automations.
• GDPR-friendly logging & analytics — scrub personal data out of text before it lands in logs, datasets or a warehouse.
• Safe sharing — clean transcripts, tickets and messages before handing them to a vendor or teammate.

Input

{
  "text": "Contact John at john@acme.com or +1 415 555 0132, card 4111 1111 1111 1111",
  "types": ["email", "phone", "creditcard", "ssn", "ip", "iban", "url"],
  "mode": "mask",
  "includeOriginal": false
}

Field	Type	Description
text	string	A single block of text to redact.
items	array	An array of strings to redact in bulk (one result row each).
types	array	Which PII to catch: email, phone, creditcard, ssn, ip, iban, url.
mode	string	mask (tokens like [EMAIL_1]) or hash (tokens with a stable hash).
includeOriginal	boolean	Include the original text in each row. Off by default.

Provide text, items, or both.

Output

Each row is one redacted input:

{
  "ok": true,
  "redacted": "Contact John at [EMAIL_1] or [PHONE_1], card [CREDITCARD_1]",
  "replacements": [
    { "type": "email", "token": "[EMAIL_1]", "value": "john@acme.com" },
    { "type": "phone", "token": "[PHONE_1]", "value": "+1 415 555 0132" },
    { "type": "creditcard", "token": "[CREDITCARD_1]", "value": "4111 1111 1111 1111" }
  ],
  "counts": { "email": 1, "phone": 1, "creditcard": 1 },
  "totalReplacements": 3
}

A combined reversible token map is also saved to the run's default key-value store under token-map.json ({ "map": { "[EMAIL_1]": "john@acme.com", ... } }) so a downstream step can restore the originals.

Output schema

Field	Type	Description
redacted	string	The text with PII replaced by tokens.
replacements	array	Each masked value: type, token and original value.
counts	object	Number of matches per PII type.
totalReplacements	integer	Total number of masked values in this item.
original	string	The original text (only when includeOriginal is on).

FAQ

Does it call an AI model or any external service? No. Detection is local pattern-matching with light validation (for example a Luhn check on card numbers) — nothing leaves the run.

How do I restore the original values? Use the token-map.json map (token → value) from the key-value store, or the per-row replacements, to swap the tokens back after your AI or automation step.

Can I automate it? Yes — via integrations on the Apify platform (n8n, Make, Zapier and more) and the Apify API.

Notes

Original clean-room implementation.