ProductsSecurity

PII redaction

Detect and mask personal information before it reaches providers or your telemetry.

Respan can detect personally identifiable information (PII) in your traffic and replace it with a placeholder. Two independent controls decide where redaction happens: on the way out to a model provider, and in the logs and traces Respan stores.

Configure it from PII redaction in your organization settings. Both modes are off by default, and changing them is an organization-level setting that can require an admin.

PII redaction settings with Gateway and Telemetry toggles, category and sensitivity controls, and a live redaction tester

Two redaction modes

ModeWhat it doesWhat it changes
Gateway (pre-redaction)Masks detected PII in the request payload before it is forwarded to the model provider.The provider never sees the raw values, so the model runs on the redacted input.
Telemetry (post-redaction)Masks detected PII in the logs and traces Respan records.Stored telemetry never contains the raw values. It does not change what the provider receives.

The two are independent. Enable Gateway to keep PII from leaving for a third-party provider, Telemetry to keep it out of your logs, or both.

Gateway redaction changes the input the model receives. If a response depends on a value that was masked, such as echoing back an order number, account for that before enabling it on production traffic.

Categories

Choose which kinds of PII to detect. The default set covers 16 categories:

names, email addresses, phone numbers, postal addresses, credit cards, Social Security numbers, bank accounts, IBANs, IP addresses, usernames, dates of birth, passports, driver’s licenses, national IDs, credentials, and demographic data.

Narrow the list to the categories that matter for your data so ordinary text is not over-redacted.

Sensitivity

Sensitivity controls how confident the detector must be before it redacts a match. Lower sensitivity redacts only clear matches and produces fewer false positives; higher sensitivity catches more borderline cases but can mask ordinary text. Tune it against real examples with the tester below.

Test before you enable

The Test panel runs your current categories and sensitivity against sample text without touching live traffic. Paste an example into Test input, and the Redacted preview shows each detected entity labeled by type, along with a count of entities detected.

Use it to confirm the settings catch what you expect and leave everything else alone before you turn redaction on.

Good to know

  • Detection is machine-learning based and probabilistic. No detector catches every case, so validate with the tester and treat redaction as a strong control rather than a guarantee.
  • Both toggles are organization-wide and apply to all traffic for the organization.
  • Redaction acts on detected entities only. Text with no detected PII passes through unchanged.