Seamless PII Redaction and Restoration

Problem

When sending user input, logs, or ticket content to an LLM, the text may contain PII such as ID numbers, phone numbers, bank card numbers, emails, detailed addresses, and names. Sending raw text increases exposure, while manual redaction is error-prone and still requires restoring placeholders in the response. This skill replaces 17 PII types before the request and restores them after the LLM returns, so callers see placeholders while end users still see the original values.

How It Works

It centers on redact.py and restore.py. Before sending, it applies rule-based matching and replaces sensitive fields with short placeholders, for example 身份证****1234, 手机****5678, 邮箱****, and [金额已脱敏]. After the response, it uses a mapping to restore the placeholders. It supports ID cards, passports, officer IDs, permits, tax numbers, mobile numbers, landlines, emails, bank cards, license plates, WeChat IDs, QQ IDs, addresses, amounts, and names. It requires Python 3.x, with 3.8+ recommended. It is suitable for LLM-facing gateways, proxies, or server-side call chains.

Boundaries

Names require context keywords such as “我”, “甲方”, or “乙方”; landlines must include the - separator; bank card numbers need preceding keywords like “银行”, “账号”, or “卡号”; detailed addresses match administrative keywords or province/city/district/street patterns. The mapping is a temporary file and should be cleaned up after the session.