PII Redaction

The redactPii option replaces personally identifiable information in the resume text with deterministic placeholders before the LLM call. After extraction, the real values are reinjected into the structured output.

The final result is identical to a non-redacted run — the LLM simply never sees the raw personal data.

Why this matters

When you send a resume to a third-party LLM API (OpenAI, DeepSeek, Anthropic), the raw text — including email addresses, phone numbers, and home addresses — leaves your infrastructure. Depending on your jurisdiction and data processing agreements, this may conflict with GDPR, CCPA, or internal data policies.

redactPii: true ensures that only anonymized text is transmitted to the LLM.

Basic usage

import { parseResume } from '@edwinfom/resume-intel'
import { createDeepSeek } from '@ai-sdk/deepseek'
 
const result = await parseResume(buffer, {
  model: createDeepSeek({ apiKey: process.env.DEEPSEEK_API_KEY })('deepseek-chat'),
  redactPii: true,
})
 
// The LLM processed "__PII_EMAIL_0__" instead of "john.doe@gmail.com"
// The output contains the real value
console.log(result.data.basics?.email) // "john.doe@gmail.com"

Works identically with streamResume():

for await (const event of streamResume(buffer, { model, redactPii: true })) {
  if (event.type === 'section') {
    // event.data already has real values reinjected
  }
}

What gets redacted

Placeholder format

Placeholders are deterministic and type-tagged:

john.doe@gmail.com  →  __PII_EMAIL_0__
contact@acme.com    →  __PII_EMAIL_1__
+1 (555) 123-4567   →  __PII_PHONE_2__
https://johndoe.dev →  __PII_URL_3__
123 Main Street     →  __PII_ADDRESS_4__

The index is a counter that increments across all redacted values in the document. This makes reinjection reliable — each placeholder maps to exactly one original value.

Advanced: custom redaction pipeline

The redaction utilities are exported for use in custom pipelines:

import { redactPii, reinjectPii, describePiiRedaction } from '@edwinfom/resume-intel'
 
// Step 1: redact
const { redactedText, placeholders } = redactPii(rawResumeText)
 
console.log(describePiiRedaction({ redactedText, placeholders }))
// "2 emails, 3 urls, 1 phone"
 
// Step 2: send redactedText to your own LLM pipeline
const extracted = await myCustomExtractor(redactedText)
 
// Step 3: reinject
const restored = reinjectPii(extracted, placeholders)

Limitations

• Unusual formats — phone numbers in non-standard formats (e.g. written out as "five five five...") are not redacted.
• Embedded in sentences — PII embedded mid-sentence without clear delimiters may not be detected.
• Custom schemas — when using outputSchema, reinjection still works but only on string fields that contain placeholders.
• OCR text — OCR output may alter the formatting of phone numbers or emails, reducing detection accuracy. In this case, redaction is best-effort.

For high-stakes GDPR compliance, combine redactPii: true with a review step on the extracted output.

Confidence scores

When redactPii is active, the confidenceScore in sectionResults reflects whether the section extracted correctly despite the redacted input:

const result = await parseResume(buffer, {
  model,
  redactPii: true,
})
 
for (const s of result.meta.sectionResults ?? []) {
  console.log(`${s.section}: confidence=${s.confidenceScore}`)
  // basics:   confidence=0.87
  // work:     confidence=1.00
  // skills:   confidence=1.00
}

A score below 0.7 on basics with redactPii: true may indicate that the redaction disrupted the extraction. In that case, try running without redactPii to compare.