Person Name Detectors

person_name

Person name detection using compromise.js Named Entity Recognition. Requires the optional compromise peer dependency.

bun add compromise

import { createRedactor, preloadPersonNameDetector } from "sensored";

await preloadPersonNameDetector();

const redactor = createRedactor({
  rules: { person_name: { action: "redact" } },
});

redactor.redact("The meeting was chaired by Dr. Sarah Johnson.");
// "The meeting was chaired by [PERSON_NAME_1]."

• ID: person_name
• Entity type: person_name
• Context required: No
• Stream supported: Yes (maxMatchLength: 40)
• Validation: compromise.js NER
• Detection reason: person_name.ner

Performance impact The person_name detector adds ~178 MiB RSS and

reduces throughput by ~4 orders of magnitude (0.06 MiB/s vs 505 MiB/s without it). Only use this when you need maximum recall and can accept the performance trade-off. :::

Not in any preset

person_name is opt-in only. It's not included in any built-in preset. You must install compromise, preload the detector, and add an explicit rule to enable it. Preloading is deduplicated, so calling preloadPersonNameDetector() more than once does not reload the dependency.

Trailing punctuation

This detector trims trailing punctuation (unless the last term is an abbreviation like "Jr.") and possessive 's.

person_name_lite

Lightweight person name detection using regex candidate discovery and a bloom filter. No runtime dependencies.

The regex finds candidate sequences (optional honorific + 1–4 capitalized words + optional suffix). Single-word names are only matched with an honorific prefix. Multi-word names require at least one word in the bloom filter AND not in the stopword set.

The bloom filter contains 99,236 unique lowercase names sourced from wikidata-names, compromise lexicon, and faker Latin locales. It uses FNV-1a + DJB2 double hashing with a measured false positive rate of ~0.1%.

const redactor = createRedactor({
  rules: { person_name_lite: { action: "redact" } },
});

redactor.redact("Send the report to John Smith.");
// "Send the report to [PERSON_NAME_1]."

• ID: person_name_lite
• Entity type: person_name_lite
• Context required: No
• Stream supported: Yes (maxMatchLength: 40)
• Validation: Bloom filter membership + stopword exclusion
• Detection reason: person_name_lite.bloom

Performance

The person_name_lite detector adds negligible overhead — no NLP dependency is required. Throughput remains at ~10 MiB/s with this detector active.

Trailing punctuation

The detector trims trailing punctuation and possessive 's from matches.