In RAG, why prompt the model to quote supporting spans before answering?

## The technique Quote-then-answer asks for a two-part response. Part one is an evidence block: verbatim spans copied out of the injected passages, each carrying the marker of the passage it came from. Part two is the answer, written only from those spans. Some teams format part one as a small structured list so it can be parsed and stripped before display; others keep it visible as the receipts panel. ## Why it works Copying is a strictly easier task than free-form synthesis. When the model must produce the supporting text first, it performs a selection step over the context that is closer to reading comprehension than to generation, and it does so before it has committed to any wording of the answer. That ordering matters: once a model has written a fluent claim, the rest of the response tends to justify it. Extracting first makes the evidence the anchor and the prose the derivative. The second reason is mechanical. A verbatim span is a string. You can test whether it occurs in the chunk it cites without another model call, which converts "is this answer grounded?" from an expensive judgment into a cheap containment check for the sourcing half of the problem. ## Formatting choices Cap the quote length. A model allowed to quote freely will sometimes emit the entire chunk, which is technically verbatim and carries no signal about which sentence mattered — a limit of one or two sentences per quote forces selection. Require the marker on each quote, not on the block. Require that the answer's claims trace to the listed quotes, and forbid new claims appearing only in the prose. If you want a machine-readable trail, ask for the evidence block in a fixed structure so a parser does not have to guess where it ends. ## Interaction with abstention Quote-then-answer pairs naturally with an insufficient-context rule: if the model cannot find a span that states the fact, there is nothing to answer from, and the empty evidence block is the trigger. That is a cleaner condition than asking the model to introspect on whether it knows enough, because the check is about the presence of text rather than a feeling of confidence. ## What it costs Output tokens roughly double for a short answer, which is a real latency and price hit on a high-volume path, and for a streaming UI it delays the first useful token unless you render the evidence block as it arrives. There is also a quality cost in some settings: heavy quoting can make answers read as stitched-together fragments, and for questions that genuinely need synthesis across five passages the evidence block gets long enough that users skip it. ## Where it fails *Correct quote, wrong conclusion.* The model quotes the vesting rule accurately and then states a consequence the rule does not imply. Grounded sourcing is not entailment; the quote block does not catch faulty inference. *Near-verbatim drift.* The model normalises whitespace, fixes a typo, expands an abbreviation or silently repairs a hyphenated line break from a PDF. The span is honest but the exact-match check fails, so normalisation and fuzzy matching are needed before you treat a mismatch as a defect. *Wrong source, right text.* Boilerplate that appears in many documents matches several chunks; the quote verifies against the cited marker by accident. *Garbage in.* If the retrieved passage is itself stale or wrong, a perfectly quoted answer is a perfectly traceable wrong answer. Quoting improves traceability, not corpus correctness — though traceability is what lets you find the bad page and fix it. ## When to use it It earns its cost where a human will check the answer: regulated advice, policy and benefits questions, contract review, anything a reviewer signs off on. It is usually overkill for casual conversational retrieval where nobody clicks through and latency is the product. A middle ground is to run quote-then-answer only for question classes flagged as high-stakes, or to keep the quotes hidden and use them purely as an internal verification input.