Research Papers Search: arXiv, PubMed, Crossref & OpenAlex API

Research papers search across arXiv, PubMed, Crossref and OpenAlex: give a topic such as "protein folding transformer" and get academic papers deduplicated by DOI, with authors, year, abstract, citation count and an open-access PDF link where one exists. PubMed articles and their abstracts come in the same list, and every row carries the DOI, the citation count and a ready BibTeX entry. Give it the titles of your references (one per entry in queries, maxResults 1) and it works as a DOI lookup that fills in the DOI and metadata of each one, so check the match before you rely on it. Sort by citations for the most cited papers, or switch on open access only. A paper row costs $1.00 per 1,000 on the free plan, and failed sources and repeats are free.

{
    "query": "protein folding transformer",
    "maxResults": 10
}

One output row per paper (shortened from the full example under Output, a real answer for another topic):

{
    "title": "Retrieval-Augmented Generation for Large Language Models: A Survey",
    "year": 2023,
    "venue": "arXiv (Cornell University)",
    "doi": "10.48550/arxiv.2312.10997",
    "pdfUrl": "https://arxiv.org/pdf/2312.10997",
    "abstract": "Large language models (LLMs) demonstrate powerful capabilities, ...",
    "citationCount": 875,
    "isOpenAccess": true,
    "sources": ["openalex", "arxiv"]
}

Research papers search: arXiv, PubMed, Crossref and OpenAlex API

Search research papers and academic papers in four official scholarly APIs at once and get one deduplicated list: OpenAlex, Crossref, arXiv and PubMed. Every paper comes with its title, authors, year, venue, DOI, link, abstract where the source allows it, citation count, an open-access PDF link and the list of sources it was found in. A one-line summaryText makes each row easy for an AI agent to quote. It costs $1.00 per 1,000 papers on the free plan and $0.60 per 1,000 on Gold and above. Failed sources and repeats are free.

Scholarly papers with the DOI, the abstract and BibTeX in every row

Tools that scrape the results page of a search engine return a snippet, a title and a citation count, and they work through proxies and CAPTCHAs. This actor reads four official scholarly APIs instead, so every row has what a reference manager and an AI agent need: the DOI, the full abstract (where the source allows it), the citation count, an open-access PDF link, the paper's type (article, preprint, review, conference paper...), a review/survey flag and a ready-to-paste BibTeX entry for an export to your reference manager.

	Scrapers of search-result pages	This actor
DOI	not returned	yes, the DOI of the published version when one is known
Abstract	a snippet	the abstract (OpenAlex, arXiv, Crossref, and PubMed when the record allows it)
Citation count	yes (the search engine's own)	yes (OpenAlex, or Crossref)
BibTeX	an extra request per paper, some scrapers	built from the row's own fields, in every row
Blocks and CAPTCHAs	common, so proxies	none: official APIs
Author profiles, h-index	some	no

It reads the public APIs of the four databases. There is no scraping, no Google Scholar and no full text of papers: only metadata, abstracts and links to open-access copies. If you need Google's own citation count or a Scholar author profile, this is not the tool. If you need the DOI, the abstract and a citation for a topic, it is.

Looking up a list of references. The actor searches by words, not by DOI. To fill in the DOIs for a list of references, put each reference title in queries (up to 50 per run) and set maxResults to 1. Each row returns the best match with its DOI, authors, year, venue and BibTeX. A match is the closest paper by title, not a proof, so compare the title and year with your reference.

What people use it for

• Literature search for an AI agent or a research assistant. Pass a topic, get papers with abstracts and citation counts as compact JSON. Each row has a summaryText and consistent fields.
• Finding the open-access PDF. pdfUrl and openAccessUrl point to free copies (arXiv, repositories, open journals). openAccessOnly returns only papers that are free to read.
• Seeing who cites what. Sort by citations to get the most cited papers first, with counts from OpenAlex or Crossref.
• Keeping up with a field. Sort by date with yearFrom to read the newest papers on a topic.
• Building a reading list across fields. A medicine paper from PubMed, the preprint from arXiv and the publisher record from Crossref come back as one row with all three in sources.

How to use it

One search:

{
    "query": "retrieval augmented generation",
    "maxResults": 10
}

Several searches, recent papers only, free to read, most cited first:

{
    "queries": ["retrieval augmented generation", "agent memory", "LLM evaluation"],
    "yearFrom": 2023,
    "openAccessOnly": true,
    "sort": "citations",
    "maxResults": 20
}

Only the preprint server and the biomedical database:

{
    "query": "CRISPR base editing",
    "sources": ["arxiv", "pubmed"]
}

A field name the actor does not know stops the run with a hint such as "maxResult" (did you mean "maxResults"?), before any source is asked and before anything is charged. A misspelt setting cannot give a run that ignored it.

Input

Field	What it does
query	One search text: keywords, a topic, a title or an author's name. Matched in title and abstract.
queries	A list of search texts, up to 50 per run. Used together with query.
sources	Any of openalex, crossref, arxiv, pubmed. Default: all four.
maxResults	Papers returned per search after the sources are merged, 1 to 200. Default 10.
sort	relevance (default), date (newest first) or citations (most cited first). OpenAlex, arXiv and PubMed sort natively; Crossref always contributes its most relevant matches, because its newest or most cited matches are mostly off topic. Everything is sorted again after the merge.
yearFrom, yearTo	Publication year range. Either can be left empty.
openAccessOnly	Only papers that are free to read. Crossref is skipped in this mode, because it has no open-access filter.
contactEmail	Optional. Sent to Crossref, OpenAlex and NCBI as the contact of their polite pools.
openAlexApiKey	Optional. A free OpenAlex key with its own daily budget (see below).
pubmedApiKey	Optional. An NCBI key: 10 requests a second instead of 3.

Output

One row per paper, in the order of the search. Fields are the same on every row (null where a source has nothing), so the CSV export works as is. An example (real data, shortened):

{
    "type": "paper",
    "query": "retrieval augmented generation",
    "rank": 1,
    "title": "Retrieval-Augmented Generation for Large Language Models: A Survey",
    "authors": ["Yunfan Gao", "Yun Xiong", "..."],
    "authorCount": 10,
    "year": 2023,
    "publicationDate": "2023-12-18",
    "venue": "arXiv (Cornell University)",
    "doi": "10.48550/arxiv.2312.10997",
    "url": "https://doi.org/10.48550/arxiv.2312.10997",
    "pdfUrl": "https://arxiv.org/pdf/2312.10997",
    "openAccessUrl": "https://arxiv.org/pdf/2312.10997",
    "abstract": "Large language models (LLMs) demonstrate powerful capabilities, ...",
    "abstractSource": "arxiv",
    "citationCount": 875,
    "isOpenAccess": true,
    "sources": ["openalex", "arxiv"],
    "paperType": "preprint",
    "isSurvey": true,
    "openalexId": "W4389984066",
    "arxivId": "2312.10997",
    "pmid": null,
    "pmcid": null,
    "bibtex": "@misc{gao2023retrieval,\n  author = {Gao, Yunfan and Xiong, Yun},\n  title = {Retrieval-Augmented Generation for Large Language Models: A Survey},\n  howpublished = {arXiv (Cornell University)},\n  year = {2023},\n  doi = {10.48550/arxiv.2312.10997},\n  eprint = {2312.10997},\n  archivePrefix = {arXiv},\n  url = {https://doi.org/10.48550/arxiv.2312.10997}\n}",
    "summaryText": "Retrieval-Augmented Generation for Large Language Models: A Survey (Yunfan Gao et al., 2023, arXiv (Cornell University)). 875 citations. Open access, PDF available. doi:10.48550/arxiv.2312.10997.",
    "retrievedAt": "2026-10-28T12:00:00.000Z"
}

• Authors are the names as the databases publish them. The row holds the first 25 names, authorCount the total. There are no e-mail addresses and no affiliation addresses.
• Deduplication. Papers are matched by DOI first, then by normalised title and year. The same paper from OpenAlex, arXiv and Crossref is one row. Every arXiv paper has a DOI (10.48550/arxiv.<id>), so a preprint meets its OpenAlex record. A preprint and its journal version have different DOIs and the same title: when the year agrees, they are one row, and the DOI of the journal version is shown.
• paperType is article, preprint, review, conference-paper, book, book-chapter, thesis, dataset or other, taken from the types the sources give (OpenAlex type, Crossref type, PubMed publication types; every arXiv record is a preprint). A paper found both as a preprint and in a journal is an article. null when no source says. isSurvey is true for a review or survey: a source calls it a review, or the title says "survey", "systematic review", "review of", "meta-analysis", "state of the art" and the like (a questionnaire survey of people is not counted).
• bibtex is a BibTeX entry built from the row's own fields: @article for articles and reviews, @inproceedings for conference papers, @misc with eprint and archivePrefix for preprints. Authors are written "Family, Given", at most 25 followed by "and others". The characters \ & % $ # _ { } ~ ^ are escaped; accented letters stay as UTF-8 (use biber, XeLaTeX or LuaLaTeX, or \usepackage[utf8]{inputenc}). Nothing is invented: volume, pages and publisher are not in the sources' search answers, so they are not in the entry. Check the entry before you cite it.
• type is paper for the rows you pay for. duplicate is a paper an earlier search of the same run already returned, and error is a source that could not answer a search. Both are free. The problems view of the dataset lists just those.
• citationCount comes from OpenAlex, or Crossref when OpenAlex does not have the paper. For a paper that only arXiv or PubMed found, the actor asks OpenAlex for its count by DOI or PubMed id, one extra request per 50 papers, so most rows have a number. The two sources count differently, so treat the number as a signal, not as a truth. It is null when no source knows the paper (a brand-new preprint), and null is not 0. If you leave OpenAlex out of sources, this extra lookup is not made.
• isOpenAccess is true when any source says the paper is free to read, false when the sources say it is not, and null when nobody knows (Crossref and PubMed say nothing about it unless a PubMed Central copy exists).

Where each field comes from, and the abstract rule

Source	Used for	Abstract
OpenAlex	title, authors, venue, year, citation count, open-access status and PDF link	Rebuilt from OpenAlex's word index.
Crossref	publisher metadata, DOIs, citation count	The abstract the publisher deposited, as plain text.
arXiv	preprints with an arXiv id, PDF link	The arXiv abstract.
PubMed	biomedical papers with PMID, journal, PubMed Central link	Only when the PubMed record carries no copyright statement.

PubMed abstracts can be copyrighted by the publisher. The actor keeps an abstract out of every PubMed record that carries a copyright statement, and the abstract of such a paper is returned only if another source (OpenAlex, arXiv or Crossref) supplies one. abstractSource says where the text came from. There is no full text of any paper in the output, only links to free copies.

Limits of the sources, said plainly

• OpenAlex. At the time of writing, OpenAlex gives callers without a key a small daily budget (about $0.10, roughly a hundred list requests) shared by everyone on the same IP address, and a free key (openalex.org/settings/api) gets its own $1 a day. Apify runs share addresses, so a run without openAlexApiKey can find the OpenAlex budget used up. The run does not fail for that: the OpenAlex part returns a free error row with the reason, and the other three sources still deliver. For regular use, add a key.
• arXiv. The arXiv API allows one request every 3 seconds, and the actor keeps to it. A run with 50 searches spends a few minutes waiting for arXiv, while the other sources work at the same time.
• Crossref and PubMed. Crossref is asked at no more than 3 requests a second, PubMed at 3 a second (10 with pubmedApiKey). If you give contactEmail, it goes to their polite pools. Without it, the actor sends none.
• A source that fails twice in a row is not asked again for the rest of the run, and the later searches get a free error row saying so. If every source fails for every search, the run fails and nothing is charged.
• Relevance is a merge of the rankings of the sources (a paper several sources rank high comes first), with OpenAlex weighted most. PubMed ranks by word matches in medicine, so for a topic outside biomedicine its hits are weak and rank low. Narrow sources to what fits your field.
• Not covered. Google Scholar, Semantic Scholar and Scopus are not used. Papers that no one has deposited with a DOI or an id in these four databases are not found.

Pricing

Pay per event: one paper event for each paper returned, $0.001 on the free plan, $0.0009 on Bronze, $0.0008 on Silver and $0.0006 on Gold and above. You pay for the papers in the output, not for what the sources found before merging and cutting to maxResults. Repeats of a paper across searches and source errors are free rows. A run stops at the cost limit you set, with the papers saved so far.

Questions

Can it fetch the full text? No. It returns metadata, abstracts and links. Open the pdfUrl of an open-access paper if you need the text.

Why is a paper missing its abstract? The source has none, or it is a PubMed record with a publisher copyright statement and no other source supplied one.

Why does the same paper appear once with two DOIs in different runs? The DOI shown is the journal version when one is known, otherwise the preprint's. A search that finds only the preprint shows the arXiv DOI.

Does it restart safely? Yes. Progress is saved after each search, and a restarted run does not save or charge the same papers again.

Did this help? A review on this page, or an issue saying what is missing, helps other people find it.