PII Protection

VaultProxy automatically detects and anonymizes Polish personally identifiable information (PII) in prompts before forwarding them to AI providers. This page documents every PII type that VaultProxy recognizes.

How It Works

VaultProxy uses a multi-layer detection pipeline:

1. HerBERT NER -- A Polish-language Named Entity Recognition model based on HerBERT (a Polish RoBERTa variant) identifies names, organizations, and locations.
2. Checksum validation -- Numeric identifiers like PESEL, NIP, and ID card numbers are validated using their built-in checksum algorithms.
3. Regex patterns -- Structured data like phone numbers, email addresses, postal codes, and IBAN numbers are matched with regular expressions.
4. Dictionary matching -- Street prefixes, city names, and other contextual clues help identify addresses and locations.

Detected PII is replaced with typed placeholders (e.g., [PESEL_1], [IMIĘ_NAZWISKO_1]) before the prompt reaches the AI provider. The original values are held in memory for up to 60 seconds during request processing and are never written to disk or logged.

Supported PII Types

Names (Imię i nazwisko)

Detection	HerBERT NER + Polish name dictionary
Placeholder	[IMIĘ_NAZWISKO_N]
Input	Proszę sprawdzić konto Jana Kowalskiego
Output	Proszę sprawdzić konto [IMIĘ_NAZWISKO_1]

Handles Polish inflected forms (Kowalski / Kowalskiego / Kowalskiemu).

PESEL

Detection	11-digit pattern + checksum validation (modulo 10)
Placeholder	[PESEL_N]
Input	PESEL pacjenta: 02271409862
Output	PESEL pacjenta: [PESEL_1]

Only valid PESEL numbers (passing checksum) are anonymized to avoid false positives.

NIP (Tax ID)

Detection	10-digit pattern + weighted checksum validation
Placeholder	[NIP_N]
Input	NIP firmy: 5261040828
Output	NIP firmy: [NIP_1]

Supports formats with and without dashes (e.g., 526-104-08-28).

REGON

Detection	9 or 14-digit pattern + checksum validation
Placeholder	[REGON_N]
Input	REGON: 012100784
Output	REGON: [REGON_1]

ID Card Number (Dowód osobisty)

Detection	3 letters + 6 digits pattern + checksum validation
Placeholder	[DOWÓD_N]
Input	Dowód: ABS123456
Output	Dowód: [DOWÓD_1]

Passport Number

Detection	2 letters + 7 digits (ICAO standard for Polish passports)
Placeholder	[PASZPORT_N]
Input	Paszport: EA1234567
Output	Paszport: [PASZPORT_1]

Phone Number

Detection	Regex for Polish phone formats (+48, 0048, 9-digit)
Placeholder	[TELEFON_N]
Input	Telefon: +48 600 123 456
Output	Telefon: [TELEFON_1]

Supports formats: +48600123456, +48 600 123 456, 600-123-456, 48 600 123 456.

Email Address

Detection	Standard email regex pattern
Placeholder	[EMAIL_N]
Input	Email: [email protected]
Output	Email: [EMAIL_1]

Address (Adres)

Detection	Polish address patterns with prefix keywords (ul., al., os., pl.) + building numbers
Placeholder	[ADRES_N]
Input	Adres: ul. Marszałkowska 10/5, 00-624 Warszawa
Output	Adres: [ADRES_1]

Recognizes common prefixes: ul., ulica, al., aleja, os., osiedle, pl., plac.

Postal Code (Kod pocztowy)

Detection	Pattern: XX-XXX (2 digits, dash, 3 digits)
Placeholder	[KOD_POCZTOWY_N]
Input	Kod: 00-624
Output	Kod: [KOD_POCZTOWY_1]

Date of Birth (Data urodzenia)

Detection	Date patterns in context of birth/age keywords
Placeholder	[DATA_URODZENIA_N]
Input	Data urodzenia: 14.02.2002
Output	Data urodzenia: [DATA_URODZENIA_1]

Supports formats: DD.MM.YYYY, DD-MM-YYYY, DD/MM/YYYY, YYYY-MM-DD.

IBAN (Bank Account)

Detection	Polish IBAN pattern: PL + 26 digits, with optional spaces
Placeholder	[IBAN_N]
Input	Konto: PL61 1090 1014 0000 0712 1981 2874
Output	Konto: [IBAN_1]

Also detects 26-digit Polish account numbers without the PL prefix.

Credit Card Number

Detection	16-digit patterns + Luhn algorithm validation
Placeholder	[KARTA_N]
Input	Karta: 4532 0151 2345 6789
Output	Karta: [KARTA_1]

Art. 9 RODO Sensitive Data

Detection	Keyword matching with Polish inflection (see Art. 9 RODO details)
Action	Flagged (not anonymized) -- returns a warning in the response
Input	Pacjent ma cukrzycę i jest leczony insuliną
Output	Request proceeds with a warning flag about sensitive data categories

warning

Art. 9 data is flagged, not anonymized. VaultProxy detects its presence and adds a warning to the response metadata. Your application should handle this flag according to your compliance requirements.

Multiple PII in a Single Prompt

When a prompt contains multiple PII items, each gets a unique numbered placeholder:

Input:

Klient Jan Kowalski (PESEL 02271409862) mieszka przy ul. Marszałkowska 10,
00-624 Warszawa. Kontakt: +48 600 123 456, [email protected].
NIP firmy: 5261040828.

After anonymization:

Klient [IMIĘ_NAZWISKO_1] (PESEL [PESEL_1]) mieszka przy [ADRES_1],
[KOD_POCZTOWY_1] Warszawa. Kontakt: [TELEFON_1], [EMAIL_1].
NIP firmy: [NIP_1].

Configuration

You can customize PII detection settings through the API:

# Get current PII settings
curl https://api.vaultproxy.ai/v1/settings/pii \
  -H "Authorization: Bearer vpx_live_YOUR_API_KEY"

# Update settings (e.g., disable specific PII types)
curl -X PATCH https://api.vaultproxy.ai/v1/settings/pii \
  -H "Authorization: Bearer vpx_live_YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "enabled_types": ["pesel", "nip", "names", "phone", "email"],
    "art9_detection": true,
    "art9_action": "flag"
  }'

Next Steps

• Art. 9 RODO -- Details on sensitive data detection
• Security -- How VaultProxy handles data securely
• API Reference -- Full endpoint documentation