Protecting Sensitive Data Before You Embed It
Once sensitive values are embedded in a vector store they are hard to remove. Tokenising them before indexing keeps raw data out of the index entirely.
Protecting Sensitive Data Before You Embed It
Vector databases are often treated like search indexes, but embeddings and their payload text can leak the very data you were trying to protect. The safest moment to deal with sensitive values is before they are embedded.
A safer pipeline
- Detect sensitive fields such as account numbers and national IDs.
- Replace them with reversible tokens.
- Store the originals in a separate, encrypted vault.
- Index only the tokenised text.
Complementary controls
- Permission filters applied inside every query, failing closed
- Per-customer schemas and namespaces
- Deletion that propagates to the vector store, and aborts if the purge fails
