tarja + Presidio

EN: two ways to get Brazilian identifiers into Presidio. PT: 2 jeitos de ter identificadores brasileiros no Presidio.

	tarja-presidio (plugin, E3.1)	upstream PR (E3.4)
EN: entities / PT: entidades	all 17 tarja entities / todas as 17	CPF + CNPJ
EN: install / PT: instalação	pip install tarja-presidio	EN: built into Presidio once merged / PT: nativo qdo aceitarem
EN: updates / PT: atualização	EN: follows tarja releases / PT: segue o tarja	EN: Presidio’s release cycle / PT: ciclo do Presidio

example / exemplo (E3.2)

from presidio_analyzer import AnalyzerEngine, RecognizerRegistry
from presidio_analyzer.nlp_engine import NlpEngineProvider
import tarja_presidio

# EN: Portuguese NLP engine (spaCy). One-off: python -m spacy download pt_core_news_md
# PT: motor NLP em portugues (spaCy). Uma vez: python -m spacy download pt_core_news_md
nlp_engine = NlpEngineProvider(
    nlp_configuration={
        "nlp_engine_name": "spacy",
        "models": [{"lang_code": "pt", "model_name": "pt_core_news_md"}],
    }
).create_engine()

registry = RecognizerRegistry(supported_languages=["pt"])
registry.load_predefined_recognizers(languages=["pt"], nlp_engine=nlp_engine)  # EN: e-mail, names... / PT: e-mail, nomes...
tarja_presidio.register(registry)  # EN: 17 Brazilian entities / PT: 17 entidades brasileiras

analyzer = AnalyzerEngine(registry=registry, nlp_engine=nlp_engine, supported_languages=["pt"])

text = "Paciente Maria, CPF 529.982.247-25, cartao SUS 898 0000 0004 3208, CEP 22290-140."
for r in analyzer.analyze(text=text, language="pt", score_threshold=0.4):
    print(r.entity_type, text[r.start : r.end], round(r.score, 2))

EN: expected: BR_CPF and BR_CNS at 1.0 (valid check digit), BR_CEP only when the context enhancer lifts it above the threshold (the word “CEP” is right before it), plus PERSON from spaCy. PT: esperado: BR_CPF e BR_CNS c/ 1.0 (DV válido), BR_CEP só se o enhancer de contexto subir acima do limite (a palavra “CEP” tá logo antes), mais PERSON do spaCy.

scores

tarja	Presidio (tarja-presidio)
EN: wrong check digit / PT: DV errado	validate_result -> False, EN: dropped / PT: descartado
EN: valid check digit / PT: DV válido	validate_result -> True, score 1.0
EN: context required (CEP, CNH, RENAVAM, PIX, CIB, IPTU, matrícula) / PT: exige contexto	EN: base 0.1, raised by Presidio’s context enhancer; use score_threshold / PT: base 0.1, sobe c/ o enhancer de contexto; use score_threshold
EN: no check digit (plate, phone) / PT: sem DV (placa, telefone)	EN: tarja’s “without context” score / PT: score “sem contexto” do tarja

EN: tarja’s own find() is stricter (context-required entities never appear without context). Use tarja directly when you don’t need the rest of Presidio. PT: o find() do próprio tarja é mais rígido (entidade q exige contexto nunca aparece sem ele). Use o tarja direto qdo não precisar do resto do Presidio.