PII Redaction¶

Lango includes a privacy interceptor that sits between the user and the AI provider. It detects and redacts personally identifiable information (PII) from user input before it reaches the LLM, preventing accidental exposure of sensitive data.

How It Works¶

graph LR
    User[User Input] --> Detector[PII Detector]
    Detector --> Regex[Regex Engine]
    Detector --> Presidio[Presidio NER]
    Regex --> Merger[Merge & Deduplicate]
    Presidio --> Merger
    Merger --> Redactor[Redactor]
    Redactor -->|"[REDACTED]"| Agent[AI Agent]

1. User input enters the security interceptor
2. The PII Detector runs all enabled detection engines (regex, Presidio)
3. Overlapping matches are merged, preferring higher-confidence results
4. Matched regions are replaced with [REDACTED]
5. The sanitized input is forwarded to the AI agent

Builtin Patterns¶

Lango ships with 13 builtin regex patterns across four categories:

Contact¶

Identity¶

Financial¶

Network¶

Pattern Customization¶

Disabling Builtin Patterns¶

Use piiDisabledPatterns to disable specific builtin patterns by name:

Settings: lango settings → Security

{
  "security": {
    "interceptor": {
      "enabled": true,
      "redactPii": true,
      "piiDisabledPatterns": [
        "ipv4",
        "kr_landline"
      ]
    }
  }
}

Adding Custom Patterns¶

Use piiCustomPatterns to add your own regex patterns:

Settings: lango settings → Security

{
  "security": {
    "interceptor": {
      "piiCustomPatterns": {
        "company_id": "\\bEMP-\\d{6}\\b",
        "internal_code": "\\bINT-[A-Z]{3}-\\d{4}\\b"
      }
    }
  }
}

Custom patterns are categorized as custom and always enabled. Each pattern must be a valid Go regex.

Presidio Integration¶

For more advanced PII detection beyond regex, Lango integrates with Microsoft Presidio -- an NER-based (Named Entity Recognition) PII detection engine.

Setup¶

Start Presidio alongside Lango using Docker Compose:

docker compose --profile presidio up -d

This starts the Presidio analyzer service on port 5002.

How It Works¶

When Presidio is enabled, Lango creates a Composite Detector that runs both engines:

1. Regex Detector -- fast, deterministic pattern matching
2. Presidio Detector -- NER-based entity recognition via HTTP API

Results are merged and deduplicated. When matches overlap, the higher-confidence result wins. Presidio scores vary by confidence level, while regex matches always have a score of 1.0.

Supported Entity Types¶

Presidio detects a wide range of entity types including:

• EMAIL_ADDRESS, PHONE_NUMBER, PERSON, LOCATION
• CREDIT_CARD, IBAN_CODE, US_BANK_NUMBER
• US_SSN, US_PASSPORT, US_DRIVER_LICENSE
• IP_ADDRESS, URL, DATE_TIME
• UK_NHS, SG_NRIC_FIN, AU_ABN, AU_TFN, IN_PAN, IN_AADHAAR
• And more (see Presidio documentation)

Graceful Degradation¶

If the Presidio service is unreachable, the detector silently falls back to regex-only detection. No errors are surfaced to the user -- the system degrades gracefully.

Configuration¶

Settings: lango settings → Security

{
  "security": {
    "interceptor": {
      "enabled": true,
      "redactPii": true,
      "presidio": {
        "enabled": true,
        "url": "http://localhost:5002",
        "scoreThreshold": 0.7,
        "language": "en"
      }
    }
  }
}

Full Configuration Reference¶

Settings: lango settings → Security

{
  "security": {
    "interceptor": {
      "enabled": true,
      "redactPii": true,
      "piiRegexPatterns": [
        "\\bCUSTOM-\\d+\\b"
      ],
      "piiDisabledPatterns": [
        "ipv4",
        "kr_bank_account"
      ],
      "piiCustomPatterns": {
        "employee_id": "\\bEMP-\\d{6}\\b"
      },
      "presidio": {
        "enabled": false,
        "url": "http://localhost:5002",
        "scoreThreshold": 0.7,
        "language": "en"
      }
    }
  }
}