PII Redaction

Privacy commitment — the load-bearing promise of this API. By definition, every request to this endpoint contains text that may include sensitive personal information. Aervik Labs does not store submitted text beyond what’s needed to process the request: no database, no logging of request bodies, and no retention of input text, output text, or detected entities once the response is returned. Our model provider sees the text: detection runs on an AI model, so each request’s text is sent to Anthropic, which processes it under its own API data-retention policy. Every request also passes through RapidAPI and Cloudflare (security). Don’t send health information: this API does not detect it.

Check warnings and unlocated_count on every response. If the model reports a value that cannot be found in your text in any written form — usually because a Unicode form or an invisible character differs (letter case, spacing and digit separators are matched) — that value is not redacted, and the response says so rather than hiding it. A value reported under a type this API does not recognize is counted the same way. Those fields report values the model found but did not redact. They cannot report a value the model never found, so a response with neither field is not proof that the text is clean. A dash-formatted SSN, a card number that passes the Luhn check, an email address written in ASCII characters, or a North American phone number written with separators is also matched by pattern, so it is redacted even when the model misses it; names, addresses and dates of birth have no such fallback. The same value written in capitals or with different spacing or separators elsewhere in the text is redacted wherever it appears.

Detected values are located deterministically in the real input text rather than trusted from the model’s own character counting, so offsets always line up exactly with the original string: original_text.slice(start, end) always equals value, with start inclusive and end exclusive. Offsets count UTF-16 code units, the way JavaScript indexes a string. Python counts most emoji as one character where JavaScript counts two, so after one text[start:end] in Python is shifted. Use redacted_text, or slice the UTF-16 encoding: text.encode("utf-16-le")[2 * start:2 * end].decode("utf-16-le"). Callers that need the raw value scrubbed from their own logs should redact it themselves before logging a raw API response.

Overlapping values become one redaction. When two values found overlap, such as a name inside an address, they are merged into one entity with one placeholder covering both. Its type is the type of the longest value, and merged_types lists each type merged, in the order the values start in the text. The same name written elsewhere in the text, outside the address, is a separate entity.

What it does not do.

• Seven types only. It does not detect IP addresses, bank account or IBAN numbers, passport, driver’s license or other national ID numbers (Social Security Numbers are U.S. only), API keys or other secrets, or health information.
• The phone pattern covers North American numbers. It matches 3-3-4 digit groups with separators, optionally with parentheses or a +1 prefix. A number written in another country’s format is redacted when the model reports it, and how often the model does has not been measured.
• One text per request, up to 50,000 UTF-16 code units. Longer text is refused with input_too_large; split it.
• No confidence scores, and one placeholder per type. Every email becomes [REDACTED_EMAIL]; placeholders are not numbered, and they change the text’s length, so offsets always index the original text, never redacted_text.
• English only, as far as we have measured. Accuracy on other languages has not been measured.
• Very dense text can keep failing. If the list of values found is too long for one model reply, the request fails with redaction_failed rather than returning a partial redaction. If the same text fails twice, split it into smaller pieces.