What is an embedding?

Definition

An embedding turns a piece of text into a long list of numbers that captures its meaning. Texts with similar meanings get similar numbers, so software can find "documents about required minimum distributions" even if they never use that exact phrase.

Embeddings are the search layer behind retrieval-augmented generation and many "ask your documents" features.

Why it matters for advisors

If a tool lets you ask questions about your CRM notes or document vault, it probably stores embeddings of that content. Those embeddings are derived from client information, so they belong in the same retention and deletion conversation as the documents themselves.

When you leave a vendor, ask whether embeddings are deleted along with the source files.

You don't need to build embeddings yourself. What you do need to know is that a vendor's search feature may store embeddings of your documents in a separate database, so ask whether those are covered by the same retention and deletion terms as the documents.

Example

A firm turns on a CRM feature that answers questions about client history. During review, the CCO confirms the vendor stores embeddings in the firm's own tenant and deletes them when a record is deleted.

Watch out for

Embeddings aren't anonymized data. Treat them like the client information they came from.