> ## Documentation Index
> Fetch the complete documentation index at: https://docs.roark.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# PII Redaction

> Mask sensitive caller information in transcripts before it reaches your team or evaluators

Roark can detect and mask personally identifiable information (PII) — phone numbers, social security numbers, credit cards, and more — in your call and chat transcripts. When redaction is on, the original PII never leaves the redaction boundary: viewers see masked tokens, evaluator LLMs receive masked text, and search results are sanitized.

Redaction is a per-project setting and applies to both **calls and chats** going forward. Existing conversations keep their original transcripts.

***

## How It Works

<Steps>
  <Step title="Transcript becomes available">
    Whether the transcript came from your integration, your own upload, or Roark's transcription pipeline, redaction kicks in once the transcript is persisted.
  </Step>

  <Step title="Roark scans the transcript for PII">
    A redaction pass identifies the entity types you've enabled (phone numbers, account numbers, addresses, names, etc.) and stores them as redaction spans alongside the transcript.
  </Step>

  <Step title="Masking on display and analysis">
    When you view a conversation, the transcript renders `[REDACTED:PHONE]`-style pills in place of the sensitive content. Evaluators that use LLMs (custom prompts, sentiment, toxicity, emotion, politeness) receive the masked transcript by default.
  </Step>
</Steps>

***

## Enabling Redaction

<Steps>
  <Step title="Open Settings">
    Go to your project's **Settings** page and find the **PII Redaction** section.
  </Step>

  <Step title="Toggle redaction on">
    Flip the master toggle. The entity-type list expands.
  </Step>

  <Step title="Choose entity types">
    Most entity types are on by default. Names and addresses are off by default — they have higher false-positive rates and customers usually opt in deliberately.
  </Step>
</Steps>

***

## Supported Entity Types

| Entity                                | Default | Notes                               |
| ------------------------------------- | ------- | ----------------------------------- |
| Social security numbers               | On      |                                     |
| Credit card number / CVV / expiration | On      | Each is a separate toggle           |
| Bank account numbers                  | On      |                                     |
| Bank routing numbers                  | On      |                                     |
| Phone numbers                         | On      |                                     |
| Email addresses                       | On      |                                     |
| Dates of birth                        | On      |                                     |
| PINs                                  | On      |                                     |
| Passwords                             | On      |                                     |
| Names                                 | Off     | Higher false-positive rate — opt-in |
| Addresses                             | Off     | Higher false-positive rate — opt-in |

***

## Evaluators and Redaction

Evaluator LLM calls (custom prompts, sentiment, toxicity, emotion, politeness) automatically receive the **redacted** transcript when redaction is enabled. This keeps PII from leaving your tenant for analysis.

If your evaluators rely on the original PII to do their job — and you accept the trade-off — you can override this with the **Allow evaluators to see original PII** toggle in the same settings panel. We recommend leaving it off unless you have a specific reason.

***

## What's Stored

* The original transcript is preserved unchanged — storage is non-destructive.
* Redaction spans (entity type, segment, confidence, source) are stored as metadata so masking is consistent across every read path and reproducible if you change the policy later.
* Audio recordings are stored as-is — audio-level redaction (beep/silence at PII timestamps) is on the roadmap.

***

## Limitations Today

* Masking is applied at the **segment** level when any PII is detected within it — so a turn containing one phone number renders as a single redaction pill, not inline word-level redaction. Word-precise inline masking is on the roadmap.
* Audio recordings are not yet redacted. If you need redacted audio (for sharing recordings outside your team), reach out — it's planned and we're prioritizing based on demand.

***

## Compliance and Boundaries

When redaction is enabled:

* The frontend renders masked tokens in transcripts and exports.
* The GraphQL API returns masked text by default for every segment-returning query (calls and chats).
* LLM evaluators receive masked text unless you explicitly opt out per project.
* OpenSearch event indexing is unaffected (event properties are structured, not transcript-derived).

If you have specific compliance requirements (HIPAA BAA, PCI scope, destructive retention of original transcripts), [reach out to support](/documentation/resources/support) — we can scope what's needed for your environment.