← All questions
System DesignHard· citation constraints· retrieval granularity· cost trade‑offs

Compliance now requires every generated claim to cite a...

Explain that the citation span and the retrieval unit need not be the same object. Index the corpus at sentence or proposition granularity for scoring, then expand each retrieved hit to its parent window before feeding it to the generator (as described in section 13.4). This expansion restores the context lost by small chunks, dissolving the recall objection. Adjust measurement: compute IoU on the expanded span (the prompt content) and evaluate the citation span separately against the annotated excerpt. Conclude with the compliance side’s cost: more vectors, an extra routing step, and added risk (section 13.3).

Code playground (Python)