piigex 0.1.0

Lightweight PII detection and redaction for structured identifiers using regex and checksum validation

piigex

A small PII detection and redaction library for structured identifiers. It uses regex plus checksum validation. There is no ML model, no NLP pipeline, and no large dependency tree.

Use it to sanitize chatbot input before it hits an LLM, scrub logs, or redact customer support transcripts. The 7 major EU countries are covered, along with international identifiers like IBAN, BIC, credit cards, email, IP addresses, and MAC addresses.

---

Quickstart

from piigex import clean, scan

clean("Send payment to ES91 2100 0418 4502 0005 1332 by Friday")
# → "Send payment to {{IBAN}} by Friday"

scan("IBAN: DE89370400440532013000")
# → [Match(name='intl_iban', token='IBAN', start=6, end=28,
#          value='DE89370400440532013000', valid=True)]

# Fine-grained control
from piigex import Scrubber
s = Scrubber(regions=["es", "intl"], stable_tokens=True)
s.clean("DNI 12345678Z and card 4111 1111 1111 1111")
# → "DNI {{ES_DNI_1}} and card {{CREDIT_CARD_1}}"

---

JSON / dict payloads

clean_json and scan_json walk nested dicts, lists, and tuples. They only touch string values; keys, numbers, booleans, and the overall structure are left untouched. scan_json returns each Match with a dotted path showing where the PII was found.

import piigex

payload = {
    "user": {"email": "[e-mail]", "iban": "GB29NWBK60161331926819"},
    "amount": 100,
}

piigex.clean_json(payload)
# → {"user": {"email": "{{EMAIL}}", "iban": "{{IBAN}}"}, "amount": 100}

for m in piigex.scan_json(payload):
    print(f"{m.path}: {m.name}")
# user.email: intl_email
# user.iban:  intl_iban

The CLI exposes the same behaviour with --json:

cat payload.json | piigex scrub --json
cat payload.json | piigex scan  --json    # match list with "path" field

---

Coverage

72 detectors across 27 regions. 58 Tier-1 identifiers are on by default. The other 14 (phone numbers and low-risk shape-only IDs) are off, since they produce more false positives. Turn them on explicitly with detectors=[...] or regions=[...].

Region	Default-on detectors
ES	es_dni, es_nie, es_cif, es_nss, es_ccc, es_referencia_catastral
IT	it_codice_fiscale, it_partita_iva
FR	fr_nir, fr_nif, fr_siren, fr_siret, fr_tva
DE	de_idnr, de_vat, de_svnr
PT	pt_nif, pt_cc, pt_niss
NL	nl_bsn, nl_btw
BE	be_nn, be_bis, be_vat, be_eid
AT	at_vnr
BG	bg_egn, bg_pnf
HR	hr_oib
CZ	cz_rc, cz_dic
DK	dk_cpr, dk_cvr
EE	ee_ik
FI	fi_hetu, fi_ytunnus
GR	gr_amka
HU	hu_anum
IE	ie_pps
LT	lt_asmens
PL	pl_pesel, pl_nip, pl_regon
RO	ro_cnp, ro_cf
SK	sk_rc
SI	si_emso, si_maticna
SE	se_personnummer, se_orgnr
intl	intl_iban, intl_eu_vat, intl_bic, intl_credit_card, intl_email, intl_ipv4, intl_ipv6, intl_mac

Countries with VAT coverage only (via intl_eu_vat): CY, LV, LU, MT.

Opt-in detectors (default disabled, feasibility="medium"):

• Phone: intl_phone_e164, es_phone, it_phone, fr_phone, de_phone, pt_phone, nl_phone, be_phone
• Shape-only IDs: es_passport, es_matricula, fr_cni, pt_passport, nl_passport, be_ogm_vcs_delimited

Enable them by name (detectors=["es_passport", ...]) or by region. They stay off by default because phone numbers and shape-only IDs are noisier.

US support is planned for v2. The internals were written to be country-agnostic, so adding new regions does not require restructuring the core.

---

Comparison

Install sizes measured with du -sh against a clean venv (baseline pip/setuptools excluded), Python 3.13. Presidio spaCy model (en_core_web_lg, downloaded separately) adds a further 560 MB.

Library	Approach	Net install size	EU structured IDs	Requires ML
piigex	regex + checksum	~6 MB	72 (58 default + 14 opt-in)	No
commonregex	regex only	~6 KB	None	No
piiregex	regex only	~4 KB	None	No
scrubadub	regex + optional NLP	~335 MB	Limited (IBAN only)	Optional
Microsoft Presidio	NLP + ML	~200 MB + model	Via custom recognizers	Yes (default)

Size breakdown for piigex: python-stdnum is 5.7 MB and the piigex source itself is about 150 KB. scrubadub 2.x pulls in scipy, numpy, scikit-learn, phonenumbers, faker, nltk, regex, and dateparser as transitive dependencies. Most of its install weight is that ML/data stack, not the detection code.

---

API reference

from piigex import Scrubber, clean, scan, Match

# Module-level convenience (uses the default Scrubber: all high/medium detectors)
scan(text: str) -> list[Match]
clean(text: str) -> str

# Configurable scrubber
s = Scrubber(
    detectors=None,            # None = default set; or ["es_dni", "intl_iban"]
    exclude=None,              # detector names to exclude from the default set
    regions=None,              # ["es", "intl"]: None = all regions
    min_feasibility="medium",  # "high" | "medium" | "low"
    validate=True,             # False = shape-only, skip checksum
    stable_tokens=False,       # True → same value → same numbered token per call
    token_format="{{{name}}}", # produces {{TOKEN}} by default
    token_map=None,            # pass a persistent TokenMap to share state across calls
)

# Match fields (frozen dataclass)
Match(name, token, start, end, value, valid, path)
#     str   str    int   int  str   bool   str   # path populated by scan_json only

Validation behaviour

validate=True (default): regex locates, checksum confirms. Matches with invalid checksums are returned with valid=False and are not replaced by clean().

validate=False: shape-match only. This is useful for catching likely PII even when the value is truncated or encoded.

Stable tokens

s = Scrubber(stable_tokens=True)
s.clean("NIE X1234567L appears again as X1234567L")
# → "NIE {{ES_NIE_1}} appears again as {{ES_NIE_1}}"

Normalized equivalents (e.g. spaced vs. compact IBAN) map to the same token index. The counter resets each clean() call unless you pass a persistent TokenMap:

from piigex.tokens import TokenMap
tm = TokenMap()
s.clean(doc1, token_map=tm)
s.clean(doc2, token_map=tm)  # same values → same indices across both docs

---

When NOT to use this library

piigex only detects structured identifiers: tax codes, account numbers, social security numbers, and other formats that have a defined shape and a checksum algorithm.

It will not detect:

• Names, organizations, or postal addresses (no NER)
• Dates of birth written in natural language
• Free-form sensitive content

If you need any of that, look at Microsoft Presidio or a spaCy-based NER pipeline. Those tools require ML models. piigex deliberately does not.

---

Documentation

• docs/index.md: landing page and design overview
• docs/coverage.md: every detector grouped by region
• docs/api.md: Python API reference
• docs/cli.md: command-line usage
• docs/extending.md: how to add a new detector

---

Scope

Structured identifiers only. No NER, no ML. EU country coverage in v1, US planned for v2. The only runtime dependency is python-stdnum.