Chat with a PDF that has actually read it. All of it, before you ask the first question.

The failure everybody has had with a chatbot and a PDF is the same one: you paste in a chapter, ask a question, and get back a confident paragraph that is about the right subject and is not in your document. It happens because the model is answering from what it learned in training, and your file was only ever context it may or may not have leaned on.

CarrelAI's chat cannot do that. Every message runs a search over the passages of your own documents first, and the answer is written from the passages that came back. There is no setting that turns this off and no mode where it answers from memory — if the search finds nothing relevant, it tells you that instead of filling the gap.

Most tools that let you chat with a PDF split it into chunks of a fixed size, or one chunk per page. Both cut arguments in half. A definition that starts at the bottom of page 4 and finishes at the top of page 5 becomes two fragments, neither of which answers a question about the definition.

CarrelAI splits along meaning instead — where the document changes subject, not where the page happens to end. It also works out the document's real structure on the way in: the actual title rather than the filename, the subject, and the chapters and sections with the pages they run over. That is why "the bit about transport" is somewhere you can jump to rather than something you have to search for.

A chat window on its own leaves you doing the real work somewhere else — the diagram gets redrawn in a notebook, the answer gets copied into a notes app, the thing you have to do about it goes on a list in a third place. Three transcriptions, three chances to lose it.

Here the chat sits beside the page it is reading, and everything it produces stays in the same workspace. Highlight any part of an answer and keep it as a note or turn it into a task with a due date. Redraw the figure on a canvas that is open next to the figure. The document reopens at the page and scroll position you left it on.

Three providers sit behind the chat — GPT-5, Claude and Gemini. You can let it choose, and if the one it chose fails mid-request the next one takes over rather than handing you an error and losing what you asked. On Pro you pick the model yourself.