security · 1 min read
PII-Aware Indexing: Keeping Sensitive Data Out of Embeddings
Embeddings can leak PII just like raw text. Here is how Bhogar AI redacts sensitive fields before indexing without breaking retrieval.
BABhogar AI TeamProduct & Engineering
Embedding a document does not anonymise it - embeddings can leak the same PII that was in the source text. For regulated industries that fact alone is a deal-breaker without proper redaction.
Why it matters
Naive redaction breaks retrieval (the embedding loses meaning). Smart redaction replaces values with type-tagged placeholders that preserve semantics while removing identity.
How Bhogar AI approaches it
Bhogar AI runs a PII detector on every chunk before embedding, replacing detected PII with type-tagged placeholders ("[EMAIL_1]", "[ORDER_ID_2]"). The original values stay in a per-tenant secret vault, retrievable only with explicit authorisation.
- Pre-embedding PII detection across 50+ entity types
- Type-tagged placeholders preserve retrieval semantics
- Originals stored in per-tenant encrypted vault
- Authorised re-hydration on retrieval
- Audit log of every re-hydration
What you get
Healthcare and financial-services customers ship RAG into regulated workflows that would have been blocked by raw-text indexing.