PII redaction
How it works#
When PII redaction is enabled for your organization, every incoming trace is scanned during ingestion. Detected sensitive values are replaced with category-specific placeholders before the data is written to storage.
The redaction runs on all text fields across your traces:
- Runs — user prompts, full prompts, responses, status messages, attributes, and metadata
- Steps — prompts, responses, status messages, and attributes
- Tool calls — arguments, results, status messages, and attributes
Nested objects and arrays within attributes and metadata are also scanned recursively.
What gets redacted#
| Data type | Placeholder | Example |
|---|---|---|
| Email addresses | [EMAIL] |
jane@example.com → [EMAIL] |
| Social Security numbers | [SSN] |
123-45-6789 → [SSN] |
| Credit card numbers | [CREDIT_CARD] |
4111 1111 1111 1111 → [CREDIT_CARD] |
| IP addresses (IPv4) | [IP_ADDRESS] |
203.0.113.42 → [IP_ADDRESS] |
| Phone numbers | [PHONE] |
+1 (415) 555-2671 → [PHONE] |
| API keys and secrets | [API_KEY] |
sk-test1234abcd5678efgh → [API_KEY] |
Credit card detection uses the Luhn algorithm to avoid false positives on arbitrary number sequences. Phone number detection requires formatting characters (parentheses, plus sign, or dash) and a valid digit count to reduce false matches.
Availability#
PII redaction is available on the Enterprise plan. Contact us at support@thecontextcompany.com to enable it for your organization.
Things to keep in mind#
- Redaction is irreversible — original values are never stored and cannot be recovered.
- Redaction is idempotent — applying it multiple times produces the same result, so placeholder tokens like
[EMAIL]are not double-redacted. - Redaction happens before data reaches storage, search indexes, and pattern detection. Downstream features like insight search and monitoring operate on the redacted text.