An answer with no source is an answer no one can check. Retrieval already knows which chunks it used, so the citation is nearly free, and requiring it changes what a wrong answer costs a team.
About 1 min of reading
Build a marked prompt.
Mark the passages
Before you run it, say out loud what it will print. Then open the answer.
The instruction
Before you run it, say out loud what it will print. Then open the answer.
Read the first example one line at a time. A line that prints shows what appears on screen; the rest set things up quietly.
A link to a file lets a reader land anywhere; a link to the passage lets them check the claim. The chunk id and its position give the second one, and that is what makes the answer reviewable.
A long answer hides which sentence came from which passage. Short answers with a citation each are checkable sentence by sentence, which is the point of the whole feature.
The retrieved chunks have ids. Put them in the prompt with a marker, ask the model to reference the marker, and map the marker back to the document. That round trip is the whole citation.
If nothing relevant was retrieved, the answer should be 'I do not have this' rather than a plausible guess. That instruction is what turns a silent fabrication into a visible gap.
Open the cited passage and read it against the answer. A citation that does not support the answer is a retrieval bug, and finding it is faster than finding the same bug in the answer's prose.
Each one is defined here on this site. Open the entry for the full definition and the other words it points at.