canpii

Canadian PII recognizers for Microsoft Presidio: provincial health numbers, passport, CRA Business Number, bank routing and driver's licences, with English and French context, LiteLLM / MCP-gateway exports and PIPEDA / PHIPA / Quebec Law 25 presets.

Why this exists

Presidio is the de-facto open-source PII detector (the presidio-analyzer package has millions of monthly downloads), but for Canada it ships only two recognizers — CA_SIN and CA_POSTAL_CODE — and both are disabled by default. Canadian health numbers, driver's licences, passports, Business Numbers and bank details are not covered, while closed products (Microsoft Purview's PIPEDA/PHIPA templates, AWS Bedrock Guardrails' CA_HEALTH_NUMBER) do cover them.

New country recognizers are merged into Presidio slowly, and that is reasonable for a project of its size. This package lets you use Canadian recognizers today, written in Presidio's own style, and we upstream what Presidio accepts.

Quickstart

pip install canpii
python -m spacy download en_core_web_sm
python -m spacy download fr_core_news_sm

from canpii import create_analyzer, get_preset

analyzer = create_analyzer(languages=["en", "fr"])  # spaCy *_sm models, threshold 0.4
text = "Carte RAMQ ZZZX 8812 2512, numéro d'entreprise 123456782 RT 0001."
for r in analyzer.analyze(text, language="fr", entities=get_preset("quebec_law25").entity_list()):
    print(r.entity_type, text[r.start : r.end], round(r.score, 2))

CA_BN 123456782 RT 0001 1.0
CA_HEALTH_NUMBER ZZZX 8812 2512 0.85

Already have an AnalyzerEngine? Add the recognizers to its registry:

from presidio_analyzer import AnalyzerEngine
from presidio_anonymizer import AnonymizerEngine
from canpii import register_all

analyzer = AnalyzerEngine()  # your existing engine (Presidio's default uses en_core_web_lg)
register_all(analyzer.registry, languages=("en",))
text = "Patient OHIP 1234-567-897 AB, BC PHN 9000 000 018."
results = analyzer.analyze(text, language="en", score_threshold=0.4)
print(AnonymizerEngine().anonymize(text=text, analyzer_results=results).text)
# Patient OHIP <CA_HEALTH_NUMBER>, BC PHN <CA_HEALTH_NUMBER>.

register_all() also enables Presidio's own CaSinRecognizer and CaPostalCodeRecognizer. Use a score threshold of about 0.4: weak, context-dependent patterns are scored well below it and only cross it when a context word ("OHIP", "carte santé", "permis de conduire", …) is nearby.

No-code / YAML setup with Presidio's RecognizerRegistryProvider:

import canpii  # registers the classes so `type: predefined` entries resolve
from presidio_analyzer.recognizer_registry import RecognizerRegistryProvider

registry = RecognizerRegistryProvider(
    conf_file=canpii.recognizer_registry_yaml()
).create_recognizer_registry()

Supported entities

Every rule, with its source URL and confidence level, is in docs/entities.md. Rules that are only secondary or community-sourced never reject a match on their own; they raise the score when they pass (pass strict=True to also reject on failure). Rules from official sources reject matches that fail them.

Other provinces and territories are tracked in docs/roadmap.md.

Gateways: LiteLLM, ContextForge and others

# LiteLLM presidio_ad_hoc_recognizers JSON (one copy per language)
canpii export --format litellm --preset pipeda -l en,fr -o canpii_ad_hoc.json
# LiteLLM config.yaml guardrail snippet (pii_entities_config for the preset, MASK)
canpii export --format litellm-config --preset pipeda
# IBM ContextForge PIIFilterPlugin custom_patterns (distinctive formats only)
canpii export --format contextforge --preset phipa
# Gateway-neutral regex + entity + score + context JSON
canpii export --format generic
# Presidio RecognizerRegistryProvider YAML (keeps full validation)
canpii export --format presidio-yaml --preset quebec_law25

Regex-only exports lose validation. LiteLLM's ad-hoc recognizers and gateway custom patterns are plain regular expressions: BC's MOD-11 checksum, the RAMQ birth-date check, the Luhn bonuses and the negative-context filter do not travel with them. Expect more false positives than with the Python package. The ContextForge export keeps only patterns with a base score ≥ 0.3, because gateways without context support would otherwise redact every 7- to 12-digit number. All exported regexes avoid look-arounds and back-references so they also compile in Rust's regex crate.

For LiteLLM, the Presidio analyzer server must have a French spaCy model loaded to use presidio_language: fr. If you can run your own analyzer image, pip install canpii into it and load the presidio-yaml export instead: you keep every check.

Policy presets

from canpii import get_preset

get_preset("phipa").entity_list()

Presets are a starting point, not legal advice — see docs/presets.md.

Benchmark

benchmarks/run_benchmark.py generates a labeled bilingual synthetic corpus (1,740 sentences: 740 positives with and without context words, 1,000 negatives including phone numbers, dates, order IDs, amounts, near-miss identifiers and "hard negatives" that put a context word next to an unrelated number) and reports precision/recall at threshold 0.4.

Positives with context words (seed 2026, presidio-analyzer 2.2.366, spaCy *_sm):

All positives, including sentences without context words:

How to read this:

• This is a regression benchmark on template sentences we wrote, not a real-world accuracy estimate. Scores and negative-context words were tuned on part of the negatives; the remaining false positives come from an "untuned" set ("Patient chart 1234567890", "Bank statement page …", "Driver ID …") where a context word sits next to an unrelated number — something no pattern recognizer can resolve.
• Context-only identifiers (Alberta PHN, BC licence, bank account, passport) are intentionally missed without context words; that is what lowers recall in the second table.
• "substring" vs "whole-word" is Presidio's context matching mode. create_analyzer() defaults to whole-word: in substring mode, upstream CA_SIN's context word "sin" matches inside "business", so every Business Number next to the word "business" is also reported as a SIN.

Reproduce with uv run python benchmarks/run_benchmark.py --show-errors.

Relationship to upstream Presidio

• This is an independent community package, not affiliated with Microsoft or the Presidio maintainers. It builds on Presidio's public APIs (PatternRecognizer, RecognizerRegistry, RecognizerRegistryProvider) and follows its contribution conventions: one recognizer class per file, COUNTRY_CODE = "ca", per-language instances, invalidate_result for checks, source URLs in docstrings.
• Credit to the Presidio maintainers and to the upstream Canadian contributors: CA_SIN, and CA_POSTAL_CODE (presidio#2204, by GitHub user watsos). We use those recognizers as they are rather than reimplementing them.
• Business Number: upstream issue presidio#2206 proposes a CA_BN recognizer and PR presidio#2207 by GitHub user Joey239716 implements it (open since 2026-07-29). Our CA_BN recognizer is an independent implementation written from the CRA documentation and deliberately uses the same entity name, so switching is seamless. Once #2207 (or an equivalent) is merged, we will defer to upstream: deprecate ours for one minor release, then delegate.
• Upstreaming plan: recognizers whose formats are backed by official sources (BC PHN, RAMQ first) will be proposed to Presidio as individual PRs, in Presidio's code style, once they have had real-world use here. See docs/roadmap.md.

Disclaimer

• Not legal advice. Presets and documentation describe engineering starting points. Compliance with PIPEDA, PHIPA, Law 25 or any other law is your responsibility.
• Detection is probabilistic. Pattern recognizers miss identifiers and flag unrelated numbers. Do not rely on this package alone to protect sensitive data.
• Synthetic data only. All identifiers in this repository, its tests and its documentation are generated or obviously fake. Randomly generated values can still coincide with issued identifiers; never use them as real identifiers. Never commit, paste or upload real identifiers — not in issues, PRs, fixtures, screenshots or logs. See SECURITY.md.
• Government programs and agencies are named only to describe the identifiers they issue; no endorsement is implied.

Contributing

Adding a province or identifier is a well-defined task — see CONTRIBUTING.md and the drafted good-first-issues in docs/issues/. Commits must be signed off (DCO).

License

Apache-2.0. Presidio is MIT-licensed; see NOTICE.