@maskera/core

Swedish-first, zero-dependency PII detection & redaction. Runs in the browser, Node, and edge runtimes. Part of maskera, live demo at maskera.dev.

pnpm add @maskera/core

import { redact } from "@maskera/core"

const { text, restore } = redact("Personnummer 19900101-2385, tel 070-174 06 58.")
text // "Personnummer [PERSONNUMMER_1], tel [TELEFON_1]."

Highlights

• Swedish-first: personnummer and samordningsnummer are date-shape checked but deliberately typo-tolerant; strict Luhn validators remain available. Organisationsnummer and payment identifiers keep checksum validation.
• Stable placeholders: the same value reuses its token within a call.
• Restore map: map LLM output back to real values, locally.
• Zero dependencies, tree-shakeable, ESM + CJS + types.

Built-in detectors

All exported individually and as defaultDetectors. Strong format checks and, where false negatives are not the larger privacy risk, checksums reject common look-alikes such as year ranges and reference numbers.

| Label | Matches | Validation | | ---------------------- | -------------------------------------- | --------------------- | | PERSONNUMMER | 19900101-2385, 900101-2385 | real date; Luhn-typo tolerant | | SAMORDNINGSNUMMER | day + 60 variant | real date; Luhn-typo tolerant | | ORGANISATIONSNUMMER | 202100-4748 | Luhn, third digit ≥ 2 | | EPOST | [email protected] | pattern | | TELEFON | 070-174 06 58, +33 6 12 34 56 78 | Swedish or explicit +CC | | POSTNUMMER | 123 45; compact 12345 only before a capitalized city or after SE- | pattern + context | | BANKGIRO | 991-2346 | Luhn check digit | | PLUSGIRO | 92 01 00-5 (spaces tolerated) | Luhn check digit | | IBAN | SE42 8000 0890 1191 4616 8423 | SE pattern | | KORTNUMMER | 13-19 digits, spaces/dashes tolerated | Luhn | | IP_ADRESS / URL | IPv4 / http(s) and www.-hosts | pattern |

Add your own with regexDetector (below); mix and match via options.detectors.

Opt-in heuristics

Three more Swedish detectors ship as heuristicDetectors (also exported individually): ADRESS (Påhittsgatan 12B), LAGENHETSNUMMER (lgh 1203) and REGNUMMER (ABC 123, currency amounts like SEK 100 excluded). They are format-based with no checksum to validate against, so they stay out of defaultDetectors; enable them when free-text addresses and plates matter more than the occasional over-redaction:

import { defaultDetectors, heuristicDetectors, redact } from "@maskera/core"

redact(text, { detectors: [...defaultDetectors, ...heuristicDetectors] })

kontonummer and journalnummer are separate contextual detectors. Ordinary Swedish bank accounts and provider-specific journal identifiers have no single checksum-backed shape, so they only match after explicit labels such as konto, kontonummer, journalnummer or journal-id. They are exported in contextualDetectors, stay out of conservative rules-only defaultDetectors, and are enabled by the hybrid pipeline.

Note: the model-hybrid redactWithNer in the maskera package enables adress, lagenhetsnummer and the context-labeled account/journal identifiers by default (its callers have free text by definition); only the synchronous redact() here keeps the structured default set.

In the 2026-08-14 end-product comparison on 258 synthetic Swedish domain texts with 952 annotated PII strings, that complete Maskera v19 hybrid fully removed 933/952 (98.0%); LogosGuard 2.4.4 in Chrome, Free/Balanced, fully removed 606/952 (63.7%). Partial/clear-text leaks were 8/11 and 49/297 respectively. The corpus is author-coupled and not exhaustively annotated for precision; see the exact settings, outputs and caveats in docs/BENCHMARKS.md.

API

redact(input, options?) => RedactResult

interface RedactOptions {
  detectors?: Detector[]                                  // default: defaultDetectors
  placeholder?: (label: string, index: number) => string // default: `[LABEL_n]`
}

interface RedactResult {
  text: string
  redactions: Redaction[]            // { start, end, value, label, replacement }
  map: Record<string, string>        // token -> original value
  restore: (text: string) => string  // bound to this call's map
}

restore(text, map) => string

Re-inserts originals into any string containing the placeholder tokens, safe on LLM output that reorders or quotes them.

regexDetector(label, globalRegex, validate?) => Detector

Build a detector. If the regex has a capture group, group 1 is treated as the value to redact. validate can reject false positives (e.g. failing a checksum).

Validators

luhnValid, isPersonnummer, isSamordningsnummer, isOrganisationsnummer.

Input size

redact() is synchronous and linear in the input size, which covers pastes, documents and book-length text. The constant factor is worst on digit-dense input: a megabyte of pure digits (a log dump) blocks the thread for roughly a second per call. For documents in the megabyte class and up, split the text on line or paragraph boundaries and redact chunk by chunk — detection is local to each chunk, so only an identifier straddling a split point is affected.

More

Part of maskera; add the maskera package for free-text names/places via the Swedish NER model. For DPOs, security teams and legal reviewers there is a whitepaper covering architecture, privacy model, training data and GDPR positioning: maskera.dev/whitepaper.pdf.

License

MIT