Scholar Search

Finds real papers through the scholarly databases themselves, not web pages about them. Every result comes from a source record with an identifier (DOI, arXiv id or PMID), so nothing here is invented. That only holds if you cite these results and nothing from memory.

• Search a topic across sources. Each work appears once, with all its identifiers, its largest known citation count, and the best abstract.
• Look up given DOIs, arXiv ids or PMIDs: the full record and an open-access link.
• BibTeX for given identifiers, straight from the DOI registries.

When to use

• "Find papers / studies / research on X", "what should I read about X", "prior work" or "related work for my paper" → search.
• "Recent preprints on X", "what's new since 2024" → --sort date or --since.
• "The most cited papers on X" → --sort citations (see workflow step 5).
• "Is there a free PDF of this paper?" → doi ID --email ...; with only a title, search it first.
• "BibTeX for these DOIs" → bibtex.

Not for:

• checking a reference list or whether a cited paper is real → cite-check;
• general web pages, news or docs → web-search;
• a paper's full text → open its oa_url with web-extract.

Quick start

The script is scripts/scholar.py relative to this skill. It uses only the standard library, so both uv run and plain python work.

uv run scripts/scholar.py "retrieval augmented generation" --source arxiv,crossref --since 2024
uv run scripts/scholar.py "base editing sickle cell" --source europepmc,biorxiv --sort date -n 20
uv run scripts/scholar.py '"graph neural networks"' --source arxiv,crossref -f markdown
uv run scripts/scholar.py "microplastics human blood"                           # four free sources
uv run scripts/scholar.py doi 10.1038/nature14539 arXiv:1706.03762 PMID:26017442
uv run scripts/scholar.py bibtex 10.1038/nature14539 1706.03762 > refs.bib

Table output (abstracts trimmed with --snippet; preprints are marked; result 1 left out here):

# retrieval augmented generation hallucination · 3 results · sorted by relevance · 2024-01-01 to now
# from arxiv 3, europepmc 3, pubmed 3, crossref 20

2. Evaluating Evidence Sensitivity and Robustness of Hallucination Detectors in Retrieval- Augmented Generation
   Ramana Kulanthaivelu · 2026-10-26 · In Review · preprint · cited 0
   doi:10.21203/rs.3.rs-11400108/v1
   https://doi.org/10.21203/rs.3.rs-11400108/v1
   Retrieval-Augmented Generation (RAG) improves the accessibility of external knowledge for large language…

3. FAIR-RAG: Faithful Adaptive Iterative Refinement for Retrieval-Augmented Generation
   Mohammad Aghajani Asl, Majid Asgari-Bidhendi, Behrooz Minaei-Bidgoli · 2025-10-25 · arXiv
   arXiv:2510.22344
   https://arxiv.org/abs/2510.22344
   open access: https://arxiv.org/pdf/2510.22344
   While Retrieval-Augmented Generation (RAG) mitigates hallucination and knowledge staleness in Large Language…

Choosing sources

--source takes a comma-separated list. The default is arxiv,europepmc,pubmed,crossref; all adds bioRxiv and any source whose key is set. Pick by field, because a source outside its field adds noise:

Query syntax

Plain words work everywhere. Each source matches them differently:

• arXiv and Europe PMC need every word. Europe PMC searches titles and abstracts.
• PubMed expands words to MeSH terms.
• Crossref ranks works by how many of the words they contain.

Quote a phrase to keep it together: '"large language models" evaluation'. A source's own syntax passes through unchanged:

• arXiv: ti:transformer AND cat:cs.CL
• Europe PMC: AUTH:"Doudna J" AND TITLE:crispr
• PubMed: crispr[ti] AND review[pt]

These fields break the other sources, so use them only with that one --source.

Options

JSON output

{"query": "...", "sort": "relevance", "since": "2024-01-01", "until": null,
 "sources": {"arxiv": 6, "europepmc": 6, "pubmed": "failed: ...", "crossref": 12},
 "count": 6,
 "results": [{"title": "...", "authors": ["Yann LeCun", "..."], "year": 2015, "date": "2015-05-27",
   "venue": "Nature", "type": "journal-article", "doi": "10.1038/nature14539", "arxiv": null,
   "pmid": "26017442", "pmcid": null, "url": "https://doi.org/10.1038/nature14539",
   "oa_url": "https://hal.science/hal-04206682", "citations": 78891, "abstract": "...", "tldr": null,
   "sources": ["crossref", "europepmc", "pubmed"], "alt_dois": [], "in_zotero": null}]}

• sources at the top level counts the hits kept from each source before merging, or says why that source failed.
• url is the DOI link when there is one. Otherwise it is the arXiv or PubMed page.
• oa_url is a free copy: the arXiv PDF, PubMed Central, Europe PMC, or Unpaywall's best location. It is null when none is known.
• citations is the largest count among the sources. It is null when only arXiv or PubMed found the work.
• alt_dois holds the work's other DOIs, for example the preprint's DOI next to the journal's.
• in_zotero is true or false with --zotero, and null without it or when Zotero cannot be reached.
• tldr is filled only by Semantic Scholar.
• doi returns {"results": [...], "not_found": [...], "failed": {id: reason}}.

Workflow for the agent

1. Choose sources by field (see the table) and put the core words of the request in the query. Use -n 10–20. Run two or three narrow queries rather than one broad one.
2. Read the abstracts before recommending a paper. A title match is not relevance. Drop results that are off topic, and say so if few remain.
3. Cite only what the script returned, with its DOI, arXiv id or PMID. Never add a paper from memory. If the user names one, look it up with doi; if they only know the title, use cite-check.
4. Say what kind each result is. Preprints (arXiv, bioRxiv, medRxiv) are not peer reviewed. A result that is also published has doi, arxiv and the journal's venue.
5. "Most cited": --sort citations ranks the matches, not the whole field, by Crossref and Europe PMC counts.
   • It works well for biomedicine.
   • For computer science it is unreliable: arXiv papers have no counts, and landmark papers can be missing. Say so, and suggest an OpenAlex key.
6. Free PDFs: pass --email with the user's real address, from the user or SCHOLAR_EMAIL. Never invent one; without it, leave --email out. Give the oa_url, and never point the user to piracy sites.
7. For a literature list, finish with -f markdown or -f bibtex. Before it goes into a paper, run it through cite-check.
8. Queries go to every source chosen. Do not search with personal or confidential details.
9. On exit 4, stop and tell the user. Do not loop: requests are already spaced.

Exit codes

One error: line on stderr with a hint after —. When some sources fail and others answer, a search still exits 0, with a note: line per failed source and the failure under sources in JSON. With doi and bibtex, ids that could be looked up are printed even when another id failed; the exit code is then that failure's.

Keys (optional)

• OpenAlex (OPENALEX_API_KEY, free at openalex.org): all fields, native citation sort. Without a key it shares a tiny daily budget per network, so it is used only with one.
• Semantic Scholar (S2_API_KEY, free on request): TL;DRs and citation counts. Without a key it answers 429 almost every time.

Keys are read from the environment and sent only to their own API, never following a redirect. Name the source in --source, or use --source all.

Zotero

--zotero asks the Zotero 7 desktop app on this computer whether each result is already in the library. It matches by DOI, including DOIs in the Extra field, or by exact title. The app must be running, with Settings → Advanced → "Allow other applications on this computer to communicate with Zotero" turned on. If it is off or closed, a note: says so and the search still works. The check is read-only, and nothing leaves the machine.

Limitations

• Metadata and abstracts only. For full text, open the oa_url. Crossref has abstracts for only some publishers.
• Ranking is each source's own. Results are interleaved by rank; ties go to works found by several sources. Crossref's relevance is the weakest: it surfaces niche works for broad queries. For --sort date|citations the script keeps only Crossref results that contain every query word.
• Citation counts differ by source (Crossref, Europe PMC, OpenAlex, Semantic Scholar count different citing sets); the largest is shown. The free sources cover biomedicine far better than other fields.
  • Free sources have no counts for arXiv-only papers, so in --sort citations those come last.
  • Classic CS papers published only on arXiv or in proceedings without DOIs can be missing from a citation ranking. An OpenAlex key helps with most of this.
• Open-access links without --email come only from arXiv, PubMed Central and Europe PMC.
• Dates are the earliest publication date a source gives, so print issues dated in the future can sort first with --sort date.
• bioRxiv and medRxiv appear a few days after posting, when Europe PMC indexes them.

Chaining

• scholar-search → cite-check before references go into a paper.
• oa_url → web-extract (or wayback if the link is dead) to read the paper.
• -f markdown → human-writing for a literature review that cites only these papers.
• -f bibtex → the user's .bib; --zotero avoids adding duplicates.