HN Explorer

33 comments

[ 7.2 ms ] story [ 84.1 ms ] thread
elliotec
This is really cool, and very relevant to something I'm working on. Would you be willing to do a quick explanation of the build?
0101111101
Sure! I first used openai embeddings on all the paper titles, abstracts and authors. When a user submits a search query, I embed the query, find the closest matching papers and return those results. Nothing too fancy involved!

I'm also maintaining a dataset of all the embeddings on kaggle if you want to use them yourself: https://www.kaggle.com/datasets/tomtum/openai-arxiv-embeddin...

heisenburgzero
So did you just combine Title+Abstracts+Authors into a single chunk and embed them or embedded them individually?
synctext
Impressive! Will you parse the papers in the future? Without citations this is not that usable for professors or scientists in general. The relevance ranking largely depends on showing these older, prominent papers. (from our lab experience building decentralised search using transformers)
cluckindan
That method can break when author names and subject matter collide.
0101111101
True, but similarly if your embeddings are any good they'll capture interesting associations between authors, topics and your search query. If you find any interesting author overlap results I'd be very interested!
bbor
Oh god, there's a medrxiv?? TIL...

Don't forget chemrXiv!

Fomite
medrxiv was very useful for keeping the various COVID-19 related preprints from completely swamping biorxiv, especially once biorxiv started aggressively rejecting them.
sitkack
embedding search via https://searchthearxiv.com/ takes either a word vector, or an abs or pdf link to an arxiv paper.

https://news.ycombinator.com/item?id=42519487

I just did a spot check, I think searchthearxiv search results are superior.

0101111101
Looks cool! You can input either a search query or a paper URL on arxiv xplorer. You can even combine paper URLs to search for combinations of ideas by putting + or - before the URL, like `+ 2501.12948 + 1712.01815`
sitkack
That is neat I like that.

It would be cool if the "More Like This" had a + button that would append the arxiv id to the search query.

0101111101
That's a nice idea! Might take a look this weekend!
masterjack
There’s also the search and browsing on https://sugaku.net, it’s more focused on math but does also have all of the arxiv on it
nblgbg
Just curious, are there any techniques other than using embeddings, computing cosine similarity, and sorting the results based on that? RRF could be used but again its very simple as well.
forrestp
My understanding is that your levers are roughly better / more diverse embeddings or computing more embeddings (embed chunks / groups / etc) + aggregating more cosine similarities / scores. More flops = better search w/ steep diminishing returns

Colbert being a good google-able application of utilizing more embeddings.

Search ends up often being a funnel of techniques. Cheap and high recall for phase 1 and ratchet up the flops and precision in subsequent passes on the previous result set.

0101111101
Exactly! A near property of the matryoshka embeddings is that you can compute a low dimension embedding similarity really fast and then refine afterwards.