Secure LLM

This system protects privacy in LLM workflows by redacting sensitive entities before any prompt is sent to the model. Detected PIIs are masked with placeholders (for example, [person_1], [email_1], and [account_1]), the sanitized prompt is then processed by the LLM, and the original values are restored in the final response.

Because the model only sees placeholders, users get nearly the same experience while private information never leaves the protected layer. The redaction pipeline covers 40+ entity types, including names, addresses, bank details, SSNs, emails, money amounts, URLs, API keys, and more.