PII Filter

Detect and redact personally identifiable information in LLM message content.

Sensitivity levels

• relaxed: email, SSN, credit card (Luhn-validated)
• balanced (default): + phone (Google libphonenumber, 250+ countries), IPv4 (octet-validated), IBAN (mod-97), DoB, salutation-prefixed names, EIN
• strict: + US passport numbers, generic hex secrets

Options: pii_types, allowlist, custom_patterns, action (redact/mask/hash), sensitivity.

Configuration

Field	Type	Default	Description
pii_types	`array	null`	None
sensitivity	string	"balanced"	'relaxed' = email, SSN, credit card only (no phone). 'balanced' (default) = + phone (libphonenumber), IPv4, IBAN, DoB, names, EIN. 'strict' = all types, no context requirement.
action	string	"redact"	Replacement format for detected PII (not credentials). 'redact' → [protected:TYPE] (default). 'mask' → ****. 'hash' → [protected:TYPE:HASH8] (referential integrity).
allowlist	array	[]	Values left untouched regardless of detection (e.g. known-safe emails).
custom_patterns	array	[]	Additional regex patterns added to the PII tier as relaxed-level detectors.

Examples

# Example 1
type: pii_filter
config:
  sensitivity: balanced

# Example 2
type: pii_filter
config:
  pii_types:
  - email
  - phone
  - ssn
  - creditcard

# Example 3
type: pii_filter
config:
  action: hash
  sensitivity: strict

# Example 4
type: pii_filter
config:
  custom_patterns:
  - name: employee_id
    pattern: \bEMP-\d{6}\b