For researchers

Read papers with something you can check. Every answer traced back to a passage you own.

For research reading, the value of an AI answer is exactly its verifiability. CarrelAI searches the passages of your own uploaded papers before every answer and writes from what it found — and on the Pro plan every citation is a link that opens the paper at that passage. You can point a question at one paper or at your entire uploaded literature, and be told which paper answered.

No card · Upload one file · The workspace builds itself

300

Documents on Pro, of up to 20,000 pages

8

Searches per question at the top plan

3

Frontier models — GPT-5, Claude, Gemini

Ask a literature, not a paper

The question that takes the longest in a literature review is rarely about one paper. It is "who else says this", "where was this measure defined", "did anyone report the opposite" — questions whose answers are spread across a folder of PDFs and are found by remembering.

The chat can be pointed at everything you have uploaded rather than at one file, and the answer comes back knowing which paper it found each part in. On Pro that reference is a link through to the passage, so the step between "it says this" and "here is the sentence" is a click.

Structure worked out on the way in

Each upload is analysed for its real title rather than its filename — which matters more than it sounds when a folder is full of files called `s41586-021-03819-2.pdf` — along with its subject and its sections with the page ranges they run over.

Passages are split along meaning rather than page boundary, so a method described across a page break is retrieved as one method. Scanned pages, which a surprising amount of older literature still is, are transcribed automatically on Study and Pro.

The reading, the working out, and the writing, in one place

The paper stays open beside a drawing canvas, which is where a model or a figure gets redrawn rather than stared at, and beside rich-text notes filed against the paper that prompted them. Highlights anchor to the words they were made on, so they survive a change of zoom or screen.

Anything in an answer can be kept in one action — as a note, or as a task with a due date carrying the name of the paper it came from. Over a long project that is the difference between a set of findings and a set of PDFs you have read once.

  • One search across every folder, document, note and task you own
  • Share a single paper by link without giving away the file or your notes
  • Answers refuse to draw on the model's own training, only your pages
  • Automatic failover between providers so a long session is not lost to an outage

What it will not do

It does not manage references, generate citations in a style, or write the paper. It is a reading and questioning environment, not a bibliography manager — if you use Zotero or similar, this sits next to it rather than replacing it.

Questions people ask before they sign up.

Can I trust the answers for research work?

The chat searches your own uploaded passages before every answer and writes from what it found, with no fallback to the model's memory — and it says so when the search finds nothing. On Pro, each citation links through to the passage so you can read the original sentence. Treat it as a fast way to find and check, not as a source in itself.

How many papers can I upload?

Three on the free plan, 50 on Study and 300 on Pro, at up to 150, 3,000 and 20,000 pages per document respectively.

Does it handle older scanned papers?

Yes, on Study and Pro. Pages that are images rather than text are transcribed automatically during the initial read, at one credit per five pages.

Which models does it use?

GPT-5, Claude and Gemini sit behind the chat. It will pick one for you and fail over to another if one is unavailable mid-request; the Pro plan lets you choose which model answers.

Upload one file and see whether it read it properly.

That is the only question worth answering before paying for anything, and the free plan exists to answer it. No card.