pii-presidio

Microsoft Presidio plugin: multi-language PII recognizers with reversible anonymization, built on pii-core and pii-veil.

Install

pip install pii-presidio
python -m spacy download pl_core_news_sm   # required for Polish NLP analysis

pii-presidio pulls in presidio-analyzer, presidio-anonymizer, pii-core, and pii-veil. spaCy itself comes via Presidio; the Polish language model has to be downloaded separately (Presidio's standard pattern).

Recognizers

from pii_presidio import get_recognizers
from presidio_analyzer import AnalyzerEngine, RecognizerRegistry
from presidio_analyzer.nlp_engine import NlpEngineProvider

nlp_engine = NlpEngineProvider(nlp_configuration={
    "nlp_engine_name": "spacy",
    "models": [{"lang_code": "pl", "model_name": "pl_core_news_sm"}],
}).create_engine()

registry = RecognizerRegistry(supported_languages=["pl"])
for r in get_recognizers(["pl"]):
    registry.add_recognizer(r)

analyzer = AnalyzerEngine(registry=registry, nlp_engine=nlp_engine, supported_languages=["pl"])
results = analyzer.analyze(text="PESEL 44051401359, email jan@example.pl", language="pl")

Each pii_core detector becomes one PatternRecognizer. Confidence scores are 0.85 for checksum-validated detectors (PESEL, NIP, REGON, IBAN, credit card) and 0.4 for regex-only ones (ID card, passport, phone, email). Per-detector context words are pre-set to common Polish keywords; pass your own via PiiCoreRecognizer(detector, context=[...]) if you need different boosts.

KRS and postal-code detectors are excluded by default (their raw regexes match ordinary 10-digit and XX-XXX strings); enable them with include_opt_in=True and pair with strict context filtering.

Reversible anonymization

from pii_veil import Mapping, Shield
from pii_presidio import ReversibleReplaceOperator, reversible_operators
from presidio_anonymizer import AnonymizerEngine

mapping = Mapping()
engine = AnonymizerEngine()
engine.add_anonymizer(ReversibleReplaceOperator)

result = engine.anonymize(
    text="PESEL 44051401359, email jan@example.pl",
    analyzer_results=results,
    operators=reversible_operators(mapping),
)
# result.text -> "PESEL [PL_PESEL_001], email [EMAIL_001]"

# Send result.text to an LLM, get a response back, then:
restored = Shield(mapping=mapping).deanonymize(llm_response_text)

The Mapping is the round-trip handle. It uses the same JSON format as standalone pii-veil, so you can interleave the two -- anonymize via Presidio, deanonymize via Shield, or vice versa.

Entity name mapping

Cross-language types use Presidio's standard names so existing pipelines that filter entities=["EMAIL_ADDRESS"] pick our recognizers up unchanged.

Sibling packages

• pii-core -- multi-language detection primitives this plugin reuses.
• pii-veil -- non-Presidio reversible anonymization with the same Mapping format.

License

Apache-2.0. See LICENSE.