unpii
High-performance French medical text anonymization library. Rust core with Python bindings.
Designed to process millions of documents efficiently. Inspired by Micropot/incognito
Installation
pip install unpii
# With Polars support
pip install unpii[polars]
Quick Start
import unpii
text = "Dr Martin au 06 12 34 56 78, email: martin@chu-brest.fr"
# Anonymize with placeholders (default)
unpii.anonymize(text)
# → "Dr <NOM> au <TELEPHONE>, email: <EMAIL>"
# Anonymize with stars
unpii.anonymize(text, style="stars")
# → "Dr ***** au *****, email: *****"
Detection Modes
Two detection levels: standard (reliable patterns) and paranoid (aggressive).
# Standard: titles, known patterns, blacklisted names
unpii.anonymize("Dr Martin est ici")
# → "Dr <NOM> est ici"
unpii.anonymize("DUPONT Jean est ici")
# → "DUPONT Jean est ici" (not detected in standard)
# Paranoid: also catches UPPERCASE Titlecase patterns, 5+ digit sequences, loose emails
unpii.anonymize("DUPONT Jean est ici", mode="paranoid")
# → "<NOM> est ici"
Custom Words to Mask
Pass additional words to mask per call. Useful when you know the patient's name:
unpii.anonymize("bob dylan est ici", mask=["bob", "dylan"])
# → "<PII> <PII> est ici"
Case-insensitive with word boundary checks:
unpii.anonymize("Bonjour Bob", mask=["bob"])
# → "Bonjour <PII>"
Ignore Groups
Skip specific categories:
unpii.anonymize("Dr Martin au 06 12 34 56 78", ignore_groups=["TELEPHONE"])
# → "Dr <NOM> au 06 12 34 56 78"
Inspect Detected Spans
Dry-run mode to see what would be masked:
for span in unpii.find_spans("Dr Martin au 06 12 34 56 78"):
print(span)
# Span(start=3, end=9, category="NOM")
# Span(start=13, end=27, category="TELEPHONE")
DataFrame Integration
anonymize_dataframe anonymizes a column in a Polars DataFrame:
import polars as pl
import unpii
df = pl.DataFrame({"text": [
"Dr Martin au 06 12 34 56 78",
"Email: joe@chu-brest.fr",
"Maladie de Parkinson",
]})
# Anonymize in place (overwrites the column)
df = unpii.anonymize_dataframe(df, "text")
# ┌─────────────────────────┐
# │ text │
# ╞═════════════════════════╡
# │ Dr <NOM> au <TELEPHONE> │
# │ Email: <EMAIL> │
# │ Maladie de Parkinson │ ← protected by whitelist
# └─────────────────────────┘
# Write to a new column
df = unpii.anonymize_dataframe(df, "text", new_column="text_anonymized")
# With options
df = unpii.anonymize_dataframe(df, "text", style="stars", mode="paranoid", ignore_groups=["TELEPHONE"])
Per-row words to mask (mask_from_columns)
Pass column names whose values are added as words to mask, per row. Useful when patient name/city are in structured columns:
df = pl.DataFrame({
"text": ["bob est ici", "alice va bien"],
"nom": ["bob", "alice"],
})
df = unpii.anonymize_dataframe(df, "text", mask_from_columns=["nom"])
# ┌─────────────────┬───────┐
# │ text ┆ nom │
# ╞═════════════════╪═══════╡
# │ <PII> est ici ┆ bob │
# │ <PII> va bien ┆ alice │
# └─────────────────┴───────┘
Global words to mask (mask)
Words to mask on every row (e.g. the doctor who wrote all reports):
df = unpii.anonymize_dataframe(df, "text", mask=["Dupont", "Cabinet Santé Plus"])
Both combined
df = unpii.anonymize_dataframe(df, "text",
mask_from_columns=["nom", "ville"],
mask=["Dupont"],
style="stars",
)
Low-level: anonymize_series
Operates on a Polars Series directly:
masked = unpii.anonymize_series(
df["text"],
mask_from_columns=[df["nom"], df["ville"]],
mask=["Dupont"],
)
df = df.with_columns(masked.alias("text_anonymized"))
Batch processing: anonymize_batch
Operates on plain Python lists (no Polars dependency):
results = unpii.anonymize_batch(["Dr Martin ici", "Email: a@b.fr"])
# → ["Dr <NOM> ici", "Email: <EMAIL>"]
Threading
Control the number of threads used by anonymize_batch, anonymize_series, and anonymize_dataframe:
unpii.set_max_threads(4) # Use 4 threads
unpii.get_max_threads() # → 4
unpii.set_max_threads(0) # Use all available cores (default)
Categories
| Group | Placeholder | Standard | Paranoid |
|---|---|---|---|
| NOM | <NOM> |
Titles + name, blacklist | UPPERCASE/Titlecase patterns, initials |
| TELEPHONE | <TELEPHONE> |
French phone numbers | — |
<EMAIL> |
Valid emails | Anything with @ |
|
| DATE | <DATE> |
DD/MM/YYYY, literal months, ISO | — |
| BIRTHDATE | <BIRTHDATE> |
né(e) le + date | — |
| ADRESSE | <ADRESSE> |
Street number + type + name | — |
| CODE_POSTAL | <CODE_POSTAL> |
5 digits + city name | — |
| NIR | <NIR> |
French social security number | — |
| IBAN | <IBAN> |
French IBAN | — |
| NUMBER | <NUMBER> |
— | 5+ consecutive digits |
| PII | <PII> |
Custom words passed via mask= |
— |
Whitelist
Medical eponyms (Parkinson, Alzheimer, Verneuil...) are protected from masking by a global whitelist, regardless of which group detected them.
API Reference
# Single text
def anonymize(text, *, style="placeholder", mode="standard", ignore_groups=None, mask=None) -> str
def find_spans(text, *, mode="standard", ignore_groups=None, mask=None) -> list[Span]
# Batch (plain Python lists, no Polars needed)
def anonymize_batch(texts, *, mask_from_columns=None, mask=None, style="placeholder", mode="standard", ignore_groups=None) -> list[str | None]
# DataFrame (requires polars)
def anonymize_dataframe(df, column, *, mask_from_columns=None, mask=None, new_column=None, style="placeholder", mode="standard", ignore_groups=None) -> DataFrame
def anonymize_series(series, *, mask_from_columns=None, mask=None, style="placeholder", mode="standard", ignore_groups=None) -> Series
# Threading
def set_max_threads(n: int) -> None # 0 = all cores (default)
def get_max_threads() -> int
# Span attributes: .start, .end, .category
Performance
Rust core with compiled regex and Aho-Corasick automata. All rules and dictionaries are embedded in the binary — zero I/O at runtime.
anonymize_batch, anonymize_series, and anonymize_dataframe use rayon for automatic parallelization across all cores.
License
MIT
See also
https://github.com/micropot/incognito https://github.com/microsoft/presidio
Metadata
Release files for unpii 0.3.1
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| unpii-0.3.1.tar.gz | 200.9 kB | Without provenance | Details for unpii-0.3.1.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| unpii-0.3.1-cp310-abi3-manylinux_2_34_x86_64.whl | CPython 3.10 | abi3 | Linux glibc 2.34+ x86-64 | Details for unpii-0.3.1-cp310-abi3-manylinux_2_34_x86_64.whl |
Total release size: 1.7 MB
Release files / unpii-0.3.1.tar.gz
| Download URL | unpii-0.3.1.tar.gz |
|---|---|
| Size | 200.9 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
adc98555c369b98a5afbce1dd2edfc7b9abdea2f768cc6b98123edbe8d4f5bc3
|
|
BLAKE2b-256 checksum How to use checksums |
b38a599247eeef8663aac4e7d07d85ce0ee1c3aa799959bb162682f237174e09
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
maturin/1.12.6
|
Release files / unpii-0.3.1-cp310-abi3-manylinux_2_34_x86_64.whl
| Download URL | unpii-0.3.1-cp310-abi3-manylinux_2_34_x86_64.whl |
|---|---|
| Size | 1.5 MB |
| Tags | CPython 3.10 Linux glibc 2.34+ x86-64 abi3 |
|
SHA-256 checksum How to use checksums |
22c3c52d7c43aaa868c1ccf247dd449e4650da35464e8b6af408af1efb0d944b
|
|
BLAKE2b-256 checksum How to use checksums |
2d74e588a1e25524a0442fa10232a5ecb7554ea5227d012b26efb52a83e58d32
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
maturin/1.12.6
|
Security for unpii 0.3.1
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.