Most "AI paper readers" hand you a summary and hide the paper. We took the opposite approach: show the real PDF, and make every sentence in it askable.

Highlight, then ask

Select any passage in the paper and a small toolbar appears:

• Ask AI — explains that exact passage in plain language and how it relates to the rest of the paper.
• Highlight — saves the passage as a private note you can return to.

Because the question carries the surrounding context, answers stay specific instead of drifting into generic description.

Everything else in one panel

On being honest about gaps

Reference data for very recent preprints is often missing from public citation graphs. When that happens the panel says so plainly rather than inventing entries — an empty state you can trust beats a list you can't.

Why the real PDF matters

Most AI reading tools show you a summary and hide the paper. That inverts the relationship: you end up trusting a paraphrase you cannot check, of a document you never opened.

Rendering the actual PDF keeps the source in front of you. The model becomes an annotation layer over the paper rather than a replacement for it, which is the only arrangement where you can catch it being wrong.

What highlight-to-ask changes

The unit of interaction becomes a passage rather than a document. "What does this sentence mean?" is a narrower and far more answerable question than "summarise this paper", and the answer can be checked against the text three centimetres away.

It also changes what you ask. Given a whole-document interface people ask for summaries. Given a selection they ask about method details, notation, and specific claims — which is what actually blocks understanding.

Notes that stay checkable

A highlight saves the quote and the page alongside your note. Three months later that is the difference between a claim you can verify and a paraphrase you no longer trust.

Summaries decay into your interpretation of your interpretation. Quotes do not.

The limits worth stating

Text extraction from PDFs is imperfect — multi-column layouts, equations and tables all degrade. Figures are not read at all, which matters because a great deal of what a paper actually demonstrates lives in its figures.

And the assistant answers from the passage you selected. It does not know the paper's wider context unless you give it, which is a real constraint on questions like "is this claim novel?"