Quoting versus paraphrasing

A grounded answer can reproduce its sources verbatim or restate them in its own words. Quoting is mechanically verifiable and often reads badly; paraphrasing reads well and introduces a place where meaning can shift. Most good systems use both, deliberately, in different places.

The choice is not stylistic — it changes what you can check.

What quoting gives you

Verifiability by string matching. A quote either appears in the cited chunk or it doesn’t. No judgement, no model call, no ambiguity. This is a large advantage and the main reason to quote.

No paraphrase drift. The words are the source’s words, so they can’t have been subtly altered.

Precise attribution. You can highlight the exact span in the original document.

A defensible record. For anything that might be questioned later, the answer contains the evidence.

What quoting costs

It reads badly. Documents are written to be read in place, with surrounding context. Lifted out, a sentence often refers to something the reader can’t see: “as described above”, “this limit”, “the aforementioned party”.

Length. Source sentences are rarely as compact as the answer needs.

Register clash. A quoted legal or technical sentence in the middle of a plain-language answer is a speed bump.

It doesn’t synthesise. Questions needing information from three places can’t be answered by quoting alone, only by stacking three quotes and leaving the reader to combine them.

Partial quotes mislead. Trimming a sentence to fit can reverse its meaning — the clause you cut was the exception.

What paraphrasing gives you

Readability. One clear sentence answering the question, in consistent register.

Synthesis. Facts from several sources combined into a coherent statement.

Appropriate length. Say exactly as much as the question needs.

Direct address. Answers the question asked rather than the question the document was written about.

What paraphrasing costs

Drift. The main risk, and it’s usually small and quiet. “May be eligible” becomes “is eligible”. “Typically processed within five days” becomes “processed in five days”. Each is a reasonable-looking compression that changes the claim.

Lost qualifiers. Conditions, exceptions and scope limits are the first things a summary drops, and frequently the most important things in the source.

Unverifiable by string matching. Checking a paraphrase needs an entailment judgement, which costs a call and has its own error rate.

False precision. A hedged source restated confidently is worse than the source, and reads as better.

The combination that works

Ask for both, use each for what it’s good at:

For each claim, first quote the exact sentence from the source
that supports it. Then give your answer in plain language.

Now you have a verifiable quote and a readable answer for every claim. String-match the quotes to validate; show the paraphrase to the user with the quote available on demand.

This is the arrangement I’d default to for any system where being wrong matters. The quote is your evidence layer, the paraphrase is your interface layer, and they’re generated together so they can’t diverge.

Where to lean each way

Lean quoting for: exact figures, limits, deadlines and thresholds; anything with legal or contractual weight; definitions as the corpus defines them; anything a user might act on directly.

Lean paraphrasing for: explanations of how something works; answers synthesising several documents; procedural summaries; anything where the source is verbose and the question is simple.

Always quote the number. If the answer contains a figure, the figure should come from a quoted span. Numbers are where drift is most damaging and most easily introduced — a paraphrase that changes “up to ten” into “ten” has changed the answer.

Instructing to prevent drift

The specific failures are predictable enough to instruct against:

Preserve hedging exactly. If a source says "may", "typically",
"up to" or "in most cases", keep that qualification.

Preserve all conditions and exceptions. Do not summarise a rule
without its exceptions.

Quote figures, dates and limits verbatim rather than restating them.

Do not make a claim more definite than the source.

Blunt, and it addresses the exact ways paraphrase goes wrong. The hedging rule alone is worth having.

Presentation

Distinguish them visually. Quotes in blockquotes or quotation marks; the answer in normal text. A reader should never be unsure which words are the source’s.

Never fabricate quotation marks. Text in quotes must be verbatim. If it’s been trimmed, mark the ellipsis; if it’s been adjusted, it isn’t a quote any more.

Progressive disclosure. Paraphrase visible, quote one click away. Most readers want the answer; the ones checking a specific claim want the evidence.

Keep the quote short. One or two sentences, with a link to the full passage. A long quote in the interface is as unhelpful as no answer.

The rule worth remembering

Quote what can be checked; paraphrase what has to be readable; and never let the paraphrase be more confident than the quote it came from.

If you generate both, the last part becomes something you can actually enforce — because the quote is sitting right there to compare against.