domherre

Redacts Swedish personal names, personnummer and e-mail addresses from string columns in a polars DataFrame.

• Personnummer — regex (date-shaped) + Luhn checksum.
• Names — KB/bert-base-swedish-cased-ner, PRS entities only.
• Medicines — names in the bundled FASS list are excluded from name redaction.
• E-mail — regex, shape-based.

Install

pip install domherre

What you get

import polars as pl
import domherre as dh

df = pl.DataFrame({
    "id": [1, 2],
    "note": [
        "Engelbert Karlsson, personnummer 850101-0006, ringde idag.",
        "Ingen känslig info här.",
    ],
})

dh.redact(df, "note")
# note -> note_redacted
# "[NAMN], personnummer [PERSONNUMMER], ringde idag."

Redacted columns are renamed with a suffix (default _redacted) so a redacted column can never be mistaken for the source. The original column is replaced, not kept. Unlisted columns are untouched. Raises if a target name already exists.

Several columns at once:

dh.redact(df, ["note", "kommentar"])

Single string:

dh.redact_text("Anna Andersson, 850101-0006")

pandas input works and comes back as pandas. A one-time PandasPerformanceWarning fires per process; silence it with DOMHERRE_QUIET=1 if you have already made your peace with pandas.

Read before you trust us

The default model is trained on SUC 3.0 — formal written Swedish. On informal text (chat logs, free-text notes, heavy abbreviation) it will miss names, silently and without error. Personnummer and e-mail are regex and deterministic; names are not.

Sample the output before calling this compliance-grade, and use scan to assert on what you are shipping rather than assuming.

Counting what was found

redact returns just the frame. redact_with_report also returns counts, and scan counts without modifying anything:

out, report = dh.redact_with_report(df, "note")
report.rows, report.rows_with_pii
report.names, report.personnummer, report.email
report.extra            # {"[TELEFON]": 3} for extra_patterns

# guard a pipeline that must never carry PII
if dh.scan(df, ["note", "kommentar"]).rows_with_pii:
    raise ValueError("PII reached a clean layer")

Ship these counts to your metrics. Without them you have no way to notice the model quietly stopped matching anything after a version bump.

Config

config = dh.Config(
    name_tag="[NAMN]",
    personnummer_tag="[PERSONNUMMER]",
    email_tag="[EMAIL]",
    suffix="_redacted",

    backend="kb-bert",       # or "gliner", "presidio", "none", or your own
    model=None,              # backend default, or a local path: "/models/ner"
    local_only=False,        # True = never hit the network, fail loudly if not cached
    num_threads=2,           # pin to your k8s cpu limit
    batch_size=32,
    min_score=0.0,           # raise to cut false-positive names, at the cost of misses
    labels=("person",),      # gliner only
    language="sv",
    backend_options={},      # passed through to the backend

    redact_names=True,
    redact_personnummer=True,
    redact_email=True,
    exclude_fass_names=True,      # preserve detected names that match the FASS list
    validate_personnummer=True,   # False = redact on shape alone, catches typo'd numbers
    enumerate_names=False,        # [NAMN_1], [NAMN_2] instead of flat [NAMN]
    extra_patterns={},            # {"[TELEFON]": r"\b0\d{1,3}[- ]?\d{5,8}\b"}
)

dh.redact(df, "note", config)

extra_patterns is applied alongside the built-ins, and wins on an identical span — so {"[EPOST]": r"\S+@\S+\.\w+"} replaces the built-in [EMAIL] tag rather than fighting it. On partial overlap the longest span wins.

enumerate_names

Distinguishes people within a single string, so downstream analysis can still tell that two mentions refer to the same person:

"[NAMN_1] mailade [NAMN_2] om [NAMN_1]s ärende"

Numbering is per string, not per DataFrame. [NAMN_1] in row 1 and row 2 are not necessarily the same person — that would need identity resolution, which this does not do.

Backend models

Name detection is pluggable. Personnummer and e-mail detection are always regex and are not affected by the backend choice. kb-bert is the default; it is the strongest Swedish-specific name model we have measured.

dh.redact(df, "note", dh.Config(backend="gliner"))

kb-bert needs the transformers extra:

pip install domherre[transformers]

gliner

config = dh.Config(
    backend="gliner",
    model="urchade/gliner_multi_pii-v1",   # default
    labels=("person", "full name"),
    min_score=0.5,
    backend_options={"map_location": "cpu", "quantize": True},
)

Custom backends

Anything with load() and person_spans(texts) -> list[list[tuple[int, int]]] works. A plain callable works too:

def detector(texts: list[str]) -> list[list[tuple[int, int]]]:
    return [[(m.start(), m.end()) for m in pattern.finditer(t)] for t in texts]

dh.redact(df, "note", dh.Config(backend=detector))

Register by name to make it selectable like a builtin:

dh.register_backend("mine", MyBackend)
dh.available_backends()   # ['gliner', 'kb-bert', 'mine', 'none', 'presidio']

Backends are cached per (name, config), so a Deployment loads each model once.

Service deployment

Load the model at startup so the first request doesn't pay for it:

dh.load(backend="kb-bert", num_threads=2, local_only=True)

Bake the weights into the image and set local_only=True. Otherwise every cold pod pulls ~440MB from HuggingFace, and a network blip becomes a runtime failure in the middle of a batch instead of a clear error at boot.

RUN python -c "from transformers import pipeline; \
    pipeline('token-classification', model='KB/bert-base-swedish-cased-ner')"

resources:
  requests: { cpu: "1", memory: "2Gi" }
  limits:   { cpu: "2", memory: "3Gi" }

Set num_threads to match the CPU limit. Torch otherwise reads the node's core count, not the cgroup limit, and thrashes threads on a large node.

Personnummer formats detected

YYMMDD-XXXX, YYMMDD+XXXX, YYMMDDXXXX, YYYYMMDD-XXXX, YYYYMMDDXXXX, plus samordningsnummer (day + 60). All must pass the Luhn check unless validate_personnummer=False.

Not detected: reservnummer/interimsnummer (region-specific, no fixed national format), and personnummer with a typo'd control digit — set validate_personnummer=False to catch those, at the cost of redacting unrelated 10 and 12-digit numbers.

License

MIT