Literature search

The query you type is parsed in your browser into structured fields — author, title, topic, collaboration, date range — which are shown as chips so you can see exactly how the sentence was understood, and corrected by hand if the guess was wrong. Each database then receives a query in its own syntax: INSPIRE's a Lisi and t "modular" and de 2023->2025, Crossref's query.author with from-pub-date filters, OpenAlex's filter= expressions.

Everything happens in your browser. This site runs no server-side code, so the requests go directly from you to the databases: nothing you type is sent to, logged by, or stored on global-nu.org.

How arXiv is reached. A browser may read a remote API only if that API sends the appropriate CORS header, and export.arxiv.org does not — a direct call from this page would be refused by the browser before you saw anything. But arXiv registers a DOI for every preprint with DataCite, under the 10.48550/arXiv.* prefix, and DataCite's API does send CORS and needs no key. So the arXiv results here are arXiv's own metadata, deposited by arXiv, fetched from the one place a browser is allowed to read it.

Ordering, and what "most cited" can honestly mean. The four sort buttons above the results work on what is already on the page: no click re-queries anything, so reordering is instant, and the order covers exactly the records the status line counts. (A fresh search does ask each database for its most recent records while the date order is active, because that genuinely changes which twenty come back.) Relevance is computed here rather than taken from any one database — once several answers are merged their own rankings are no longer comparable — from the authors and the title and topic words you asked for, a collaboration name if you gave one, and the number of databases that independently returned the paper. Its citations and its age are a separate, second number, read only to separate two papers the query itself cannot tell apart — never added to the match, so a much-cited paper cannot out-rank one that answers your query better. The full weighting is written out above relevance() in the page's script. Sorting by citations is a button of its own, and it is deliberately not pushed upstream: asking INSPIRE for sort=mostcited does not return the most cited papers on a subject, it returns the most cited papers that mention the words — which is how a search for lecture notes on string theory once came back with the PDG Review.

When INSPIRE says nothing. INSPIRE does not degrade when a question is slightly wrong — it answers zero, which looks exactly like "no such paper". So this page never takes one silence for an answer. A momentary failure is retried once. A name is read in every spelling it might have: a particle stays with its surname (de Salas, van der Meer), an initial stays with its name (E. Lisi), and two bare words are sent as both readings at once, since lisi marrone means two authors and antonio marrone means one person and nothing in the two strings tells them apart. An arXiv identifier or a DOI, pasted on its own, is asked of INSPIRE as an identifier rather than as words. And when a question still comes back empty, the constraints come off one at a time — the dates, then the last name, then the topic words, then the title — until something answers; if a looser question was the one that answered, the line above the results says which constraint had to go.

Where the citation counts come from. INSPIRE, and nowhere else. Every database here counts citations and they do not agree, because they index different literature; mixed into one column those numbers stop being comparable and ranking by them would reward the paper indexed by the most generous counter. So a record INSPIRE returned carries INSPIRE's number whichever database found it first, and a record INSPIRE did not return is looked up by its arXiv identifier or DOI — INSPIRE answers twenty results at a time, and a paper it knows perfectly well is often simply not among the twenty it sent back. What no lookup finds is shown with no count at all rather than borrowing one, and sorts as unknown. Untick INSPIRE above and no counts appear at all: you have asked this page not to use the one source it trusts for them.

How the subfields are decided. Each result is filed under one subfield, and the grouping is not invented here: it is the arXiv subject taxonomy — the classification under which these papers were actually deposited — collapsed into eleven groups. Whatever each database holds of it is used and merged: INSPIRE's arXiv categories and its own subject terms, arXiv's own subject headings as deposited with DataCite, OpenAlex's topic, Semantic Scholar's fields of study, Crossref's subject when a publisher deposited one. Neutrino physics is the one group that is not an arXiv category — it runs through hep-ph, hep-ex, astro-ph.HE and nucl-ex at once — so it is decided from the topic, or from the words of the title or of the journal or conference it appeared in, and it takes precedence over the rest. A record no database classified is shown as Unclassified, not guessed into a group.

Why not NASA ADS. Its search API is public and documented, but it requires a personal token, which would have to be published in the page source to work — so for ADS the parsed query is turned into a correct search URL and offered as a button instead.

Results are capped at 20 per database, and every outbound link opens in a new tab, so you never lose your place here.