1. What it is

The chat at the top of my home page is Ask My Papers, which I build at Tasman Lab. A researcher adds their papers and gets an assistant that answers questions from those papers and nothing else. Each answer points to the paper and the PDF page it came from.

I wanted it because people ask me about my work in places where I can't hand them a PDF: after a talk, in an email, on my website. Most of them want one finding or one number, not the whole paper. An assistant that can find that passage and show where it sits saves both of us time, as long as it never makes things up.

2. Reading papers properly

Everything depends on reading the paper well in the first place. My first attempt treated a paper as plain text. That was mostly fine for my epidemiology papers, then a mathematician tried it. Equations came out as scattered symbols, tables became runs of numbers, and nobody could tell which section a paragraph belonged to. Questions about a theorem or a model's parameters got vague answers.

Ask My Papers now keeps that structure. Equations stay equations, tables keep their rows and columns, and every passage remembers its section and its page. When an answer needs maths, the maths is shown properly rather than as a string of symbols.

3. Finding the right passages

A question rarely uses the same words as the paper that answers it. A visitor might ask how a virus survives between outbreaks when the paper talks about chronic infection. Another might type the exact name of a gene or a method, and expect that precise term to be found. The assistant handles both kinds of question.

Many good questions need more than one paper. “How do these two studies differ in their approach?” is only useful if both studies are heard from. The assistant can draw on several papers at once for a single answer, and it tries not to pull in papers that merely share the topic.

4. Answering only from the papers

The answer is written from the passages it found in the researcher's own papers. It does not fill gaps from general knowledge, and it does not guess.

When the papers don't cover a question, it says so plainly: “These papers don’t cover that.” I care about this more than anything else in the product. A researcher's name sits next to these answers, and a confident wrong answer does them real harm. A clear “not covered” is a good outcome.

5. Follow-up questions

People rarely stop at one question. After an answer you can ask “and what did the other paper find?” or “what was the sample size?” and the assistant reads it in the light of the earlier answer. When you want to change subject, a New topic option starts afresh, so an old thread doesn't colour the next question.

6. Citations you can check

Each answer lists the papers it drew on, with the PDF page for each one. That list comes from the passages that were actually used, so it can only name papers in the researcher's own collection, and the page is the real page.

I think of the citations as the point of the whole thing. You shouldn't have to trust the answer. You should be able to open the paper at that page and read it for yourself.

7. Privacy

A researcher can also keep a private collection, for the papers they read rather than write. Only they can see it or ask it, and it is never searched for anyone else. Papers, public or private, are never used to train a model.

8. Testing before release

“It gave a good answer when I tried it” is not evidence. Before any change is released, the assistant is run against a fixed set of test questions and the answers are checked.

The sets are small, so I don't quote an accuracy figure from them. What they do well is catch a change that makes things worse before a visitor sees it.

9. Try it

Ask my assistant something my papers cover, then something they don't, and compare the two answers. Check a page number or two while you are there. If you are a researcher, you can set one up for your own papers.

See it on my papers, or find out more about Ask My Papers.

  Try it on my papers   Get one for your papers