Retrieval · Governance

Citations are the audit trail

12 October 20264 min readZegress.Ai

When an AI assistant gives an answer in a regulated business, someone may later have to defend it. A compliance officer explaining a decision to a regulator needs to show where the answer came from, and so does an investigator building a case file. A citation is how they do that.

An answer without a source is an opinion

Most assistants can produce fluent answers from documents they have been given. Fewer can show, reliably and every time, which file and which page each fact came from. Without that, the person relying on the answer has to choose between trusting it blind and checking everything by hand. Neither scales.

What a useful citation looks like

In K-Lake, every answer links back to the exact source document, opening at the page, or for a recording at the timestamp, where the fact appears. Permissions are checked again when the link is opened, so a citation never becomes a back door to a file the reader cannot see.

The citation and the question are both recorded, which means the evidence for an answer survives after the conversation ends.

Why retrieval quality decides this

A citation is only as good as the passage behind it. K-Lake combines keyword and meaning-based search, so it finds the passage that answers the question rather than the one that repeats its words. In internal benchmarking that found 62% more correct answers than keyword search alone.

What to ask any vendor

Ask for a sample answer and follow every citation to its source. Ask what happens when the reader no longer has access to the cited file, and whether the question and its sources are logged, and for how long. The answers tell you whether citations are built into the product or added to the interface.

Zegress is the Zero Egress AI Platform. The document estate, index and permission enforcement stay inside your perimeter. See it on your own data, or read the trust centre if you are the one who has to sign it off.