Named Entity Recognition

Extract named entities — people, organizations, locations, and miscellaneous entities — from text using token classification models.

For full API reference (extractEntities(), extractEntitiesMany(), options, result types, and custom providers), see the Core Classification guide.

See it in action

Try the PII Redactor block — it extracts named entities to redact PII from documents, on-device.

Recommended Models

Model	Size	Entity Types	Use Case
`Xenova/bert-base-NER`	~110MB	PER, ORG, LOC, MISC	General-purpose NER

Entity Types

Type	Description	Examples
`PER`	Person names	Tim Cook, Marie Curie
`ORG`	Organizations	Apple, United Nations
`LOC`	Locations	Cupertino, France
`MISC`	Miscellaneous	English, Nobel Prize

Document Redaction Example

Based on the PII Redactor block:

import { transformers } from '@localmode/transformers';
import { extractEntities } from '@localmode/core';

const model = transformers.ner('Xenova/bert-base-NER');

const { entities } = await extractEntities({
  model,
  text: document,
  abortSignal: controller.signal,
});

// Strip BIO prefix from entity types (B-PER → PER, I-ORG → ORG)
const cleaned = entities.map((e) => ({
  ...e,
  type: e.type.replace(/^[BI]-/, ''),
}));

// Redact entities by replacing with placeholder
let redacted = document;
for (const entity of cleaned.reverse()) {
  redacted =
    redacted.slice(0, entity.start) +
    `[${entity.type}]` +
    redacted.slice(entity.end);
}

Best Practices

NER Tips

Handle BIO prefixes — Raw output uses B-/I- prefixes (B-PER, I-PER); strip them for display
Use offsets for redaction — start/end give exact character positions for replacement
Process in reverse — When redacting, work from end to start to preserve offsets
Filter by score — Low-confidence entities may be false positives