Ask your library: grounded chat with a citation trail

Ask a general-purpose chatbot about a paper and you get one of two things: a confident answer stitched together from whatever the model happened to absorb in training, or an honest "I don't have access to that." Neither is what you want mid-literature-review. You want an answer that comes from your papers, and a way to check it.

That's what ScholarThread's chat is for. It's not a chatbot bolted onto the side of your library — it's a search-and-answer layer over the papers you've actually imported, and every claim it makes points back to the passage that supports it.

What it actually does

Ask a question and ScholarThread searches across your whole library — not just whichever paper you happen to have open — pulls the passages relevant to your question, and drafts an answer grounded in that text. Claims in the answer carry a citation, rendered as a small numbered chip. Click one and you land on the exact highlighted passage in the source paper, in the reader, not just the paper's title page.

That round trip — question, grounded answer, click-through to the sentence that backs it up — is the whole point. An answer you can't verify is just a plausible-sounding guess with better formatting. An answer with a citation trail is something you can actually build an argument on.

A cross-paper question, not a single-paper one

The most useful questions aren't "summarize this paper" — you can just read the paper for that. They're the ones that span your library: do my papers on retrieval-augmented generation agree on where the privacy risks actually are? Which of these five papers make the strongest case against the assumption in my second chapter? That's the kind of synthesis a single-paper reader can't do and a general chatbot can't do either, because it doesn't have your library.

Grounded chat pulls from across everything you've imported, so an answer can (and often does) cite passages from several different papers in the same response — each one a click away from where it actually says that.

Or narrow it to one paper

Sometimes the question really is about a single paper — you just want it answered without reading the whole thing first, or you want to interrogate a specific claim in detail. Open a paper and start a chat scoped to it, and the assistant answers from that paper alone rather than searching your whole library. The citation trail works exactly the same way: every claim points back to the passage in this paper, one click away.

Each conversation is its own session, so a scoped chat about one paper and a broader question about your whole library don't tangle into a single thread — you can hold several running conversations at once and come back to any of them later with its history intact.

What it needs from your library first

Chat can only answer from text it has actually indexed. A paper that's still processing, or one whose PDF only ever resolved to an abstract, gives the assistant less to work with — and it'll tell you that directly rather than guess past the gap. If an answer feels thinner than you expected on a paper you know has more to say, check that paper's indexing status before assuming the assistant missed something.

It also can't add papers to your library on your behalf. If you ask about work you haven't imported yet, expect it to say so rather than answer from something it read about your topic in general. Import it — via Zotero sync, OpenAlex, or a direct upload — and ask again.

Doing this in ScholarThread

Open Chat, ask a question the way you'd ask a colleague who'd actually read your library, and read the answer with its citation chips. Click one to jump straight to the passage it came from, check it against the source the way you'd check any claim, and keep going from there — ask a follow-up, ask it to compare two papers, or ask what a passage means for a specific chapter of your own draft.

It's currently a Pro feature. The point isn't to replace reading — it's to make the first pass across a library of a hundred papers take an afternoon instead of a week, with a way to verify every claim it hands you.