Reading a piece of RAG

一月 12, 2026

A few days ago, someone submitted to Hacker News a link to sciencestack.ai, but received very little fanfare.

2 points by cjlooi 4 days ago

But what the product is trying to do is of much interest to me. As an editor, I have imagined a future for publishing where authors write in Markdown or some more semantic markup languages so more interactive ways of content consumption can build on them. And as a product manager working on a scholar AI agent product, I want to see how people approach the problem of how to use agents to facilitate reading and understanding of science papers.

It is a bit hard to believe, given the huge progress LLM products achieved in 2025, that LLMs still don't read PDFs semantically. They can extract text alright, even fixing some problems like line breaks along the way, but they don't know easily where to find "Figure 3.2", or which line is a section heading, as we would imagine. So sciencestack chooses to only process papers with their LaTeX code available, which are almost exclusively on ArXiv as of now.

By parsing LaTeX codes, you get much more semantic and structural information on the text than by looking at (text extracted from) PDFs. So sciencestack can provide an API where you can query a paper as if it is a web service, and make AI agents who can answer your reading questions better.

They combined these into an interactive paper reader. The user can see immediately what "Section 3.2" or "Figure 5" or "Reference [2]" is, by hovering the cursor over the highlighted/linked text. Also, thanks to the LaTeX source code, they can copy math equations in one click.

For each section, LLM (GPT-5 Nano) generates a summary. The user can view them as a list in a sidebar, or click the annotation icon in the right margin, just like in MS Word. But they cannot interact with these annotations further, like commenting etc. In this case, I think it is better if these notes are integrated into the table of contents to the left.

If you select some text in the paper, you get a popup toolbar which lets you copy LaTeX/Markdown presentation of the text selected, add annotation, or ask AI about it, with the latter two features only available to signed-in users.

By the way, when I select the text, there are now 3 popup toolbars offering to do something with the selection with LLMs. They are respectively provided by the instant message app, the browser, and sciencestack. This reminds me of late 1990s when all kinds of dashboards and toolbars are competing for your desktop computing and web surfing experiences.

Sciencestack is more like a prototype or proof of concept, than a product that convince you that a new paradigm is on the horizon. It also reminds me that the intelligent reader I want to build may be still too computationally expensive/difficult to be feasible. At the current state, it may help one read a paper better, or more conveniently extract data, but the improvement is very limited, not enough to justify an independent product and all those tokens you have to buy.