For developers

Install maskera, create a recogniser once and mask text locally in the browser or Node. The Swedish AI model is downloaded on first use and then reused from the local cache.

Install

npm install maskera @huggingface/transformers

Rules and AI model are included; import the entire API from maskera.

Mask text

import { createNerRecognizer, redactWithNer, type NerRecognizer } from "maskera"// maskera's Swedish model, about 43 MB, runs locallyconst recognizer: NerRecognizer = createNerRecognizer()const { text, restore } = await redactWithNer(  "hej jag heter anna karlsson, personnummer 19900101-2385, och bor i uppsala",  { recognizer },)text// "hej jag heter [NAMN_1], personnummer [PERSONNUMMER_1], och bor i [PLATS_1]"

Diagram of maskera's two layers: your text is split between rules for fixed-format data and a Swedish AI model for free text, then merged into masked text. — Rules handle fixed-format data. The AI model handles free text such as names and addresses. Rules win when detections overlap. The model categories and default rule-engine detectors are listed category by category. See exactly what is masked.Open full-size diagram

Send to the AI service and restore the response

// send the masked text to any AI service// detected personal data has been replaced with placeholdersconst answer: string = await fetch("https://api.example.com/chat", {  method: "POST",  body: JSON.stringify({ prompt: text }),}).then((r) => r.text())restore(answer)// placeholders are replaced with the originals, locally

Optional profile for clinical text

The default mode suits mixed text. For medical records and clinical workflows, a named profile protects clinical facts while rule-based personal data is still always masked. Read more about the clinical profile on GitHub

JS / TS

const { text, restore } = await redactWithNer(journalText, {  recognizer,  profile: "clinical",})