| name | semantic-scholar |
| description | Search academic papers, or list one paper's references. Returns abstracts plus metadata (including a pdf_url) — use fetch-text on that pdf_url to read the paper itself. Use for scholarly literature, not general web content. |
| args | {"query":"required string (unless paper_id given) — search terms (e.g. \"constitutional AI\", \"BERT pretraining\")","paper_id":"optional string — return this paper's reference list instead of searching. Accepts an S2 paperId, \"arXiv:1706.03762\", or \"DOI:10.1145/3442188\"","limit":"optional int — max results for a search (default 10), or max references returned (default 40)"} |
semantic-scholar
Search the Semantic Scholar academic corpus, or pull the structured
reference list of a single paper.
Two modes
- Search (
query) — returns each paper's abstract plus metadata:
title, authors, year, venue, citation count, DOI, paper_id, and
pdf_url. Abstracts only; this call does not read PDFs.
- References (
paper_id) — returns that paper's reference list as
resolved records: title, year, authors, venue, and an arXiv id or DOI
where one exists. These come from the citation graph, so each entry is
a real paper you can look up directly rather than a string scraped
from a bibliography.
Reading the actual paper
Search gives you a pdf_url. Pass it to fetch-text to read the paper.
For a research PDF, fetch-text returns a section index first, so you
can then request the one or two sections you actually need instead of
pulling tens of thousands of words into context.
When to use vs search-web
semantic-scholar — scholarly literature, peer-reviewed work, papers, citations.search-web — current events, general web content, anything that isn't academic.
For a comprehensive search across both, do both — they cover different corpora.
Examples
{"thought": "find recent constitutional AI papers", "tool": "semantic-scholar", "query": "constitutional AI alignment"}
{"thought": "what does the transformer paper build on?", "tool": "semantic-scholar", "paper_id": "arXiv:1706.03762"}
{"thought": "load BERT-architecture papers, just the top three", "tool": "semantic-scholar", "query": "BERT bidirectional transformer pretraining", "limit": 3}
Notes
- Free public API; no key required. Set
SEMANTIC_SCHOLAR_API_KEY for
higher rate limits — unauthenticated use throttles quickly (429), and
a throttled call reports itself as unavailable rather than as
"no results".
- Papers with neither an abstract nor a PDF are dropped from search results.
- A paper S2 has no record of has no references here; extract them from
the PDF itself in that case.
