Tag every part of a generated sentence that came from a source, merging overlaps and listing the sources that contributed.

This is the coding half of a product every RAG deployment eventually needs, which is why it is asked: the interviewer wants to see whether you recognize interval merging under an unfamiliar description. Two clarifying questions are worth real points. Do overlapping occurrences of the SAME source count once or twice, since that decides the ordering of the attribution list. And do touching spans merge, since 'overlapping' is ambiguous at the boundary and choosing silently is how two correct-looking implementations disagree.