API

anonymize(text, lang=None, *, keep=(), detectors=(), protect_authorities=True)

Returns (text_with_placeholders, mapping).

• lang: "de", "fr", "it", "en" or None (a region such as "de-CH" is accepted). Anything else raises ValueError. It has no effect: all four languages are always active, because Swiss texts mix them.
• keep: values that stay in the text: exact strings (case-insensitive) or compiled patterns that must match the whole value.
• detectors: extra recognisers, for example presidio_detector(...) from isomorphism.presidio.
• protect_authorities: names such as "Steueramt St. Gallen" are not replaced (default True).

Identical values get the same placeholder inside one call. A placeholder that already stands in the input text, like a literal [AHV_1], is never reused.

restore(text, mapping)

Replaces placeholders with the real values in one pass. Models sometimes change the case or add spaces inside the brackets ([ahv_1], [ AHV_1 ]); both still match. Placeholders that are not in the mapping stay unchanged. mapping can be any dict[str, str], for example one you loaded from your own storage.

detect(text, lang=None, *, keep=(), detectors=(), protect_authorities=True)

Returns list[Finding] without changing the text.

restore_stream(chunks, mapping) and Restorer(mapping)

restore_stream takes an iterable of text pieces and yields restored pieces. Restorer.feed(piece) does the same step by step; call Restorer.flush() after the last piece. An end that could be the start of a placeholder ([AHV_) is held back until the next piece or the flush, so a placeholder split between two pieces is restored. At most 40 characters are held back.

Finding

A Finding holds no value.

Mapping

A dict from placeholder to real value with a findings list. repr() and str() show only the number of placeholders, so the values do not end up in logs by accident. The package does not log anything.