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.
1---2name: semantic-scholar3description: 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.4---56# semantic-scholar78Search the Semantic Scholar academic corpus, or pull the structured9reference list of a single paper.1011## Two modes1213- **Search** (`query`) — returns each paper's **abstract** plus metadata:14 title, authors, year, venue, citation count, DOI, `paper_id`, and15 `pdf_url`. Abstracts only; this call does not read PDFs.16- **References** (`paper_id`) — returns that paper's reference list as17 resolved records: title, year, authors, venue, and an arXiv id or DOI18 where one exists. These come from the citation graph, so each entry is19 a real paper you can look up directly rather than a string scraped20 from a bibliography.2122## Reading the actual paper2324Search gives you a `pdf_url`. Pass it to `fetch-text` to read the paper.25For a research PDF, `fetch-text` returns a section index first, so you26can then request the one or two sections you actually need instead of27pulling tens of thousands of words into context.2829## When to use vs `search-web`3031- `semantic-scholar` — scholarly literature, peer-reviewed work, papers, citations.32- `search-web` — current events, general web content, anything that isn't academic.3334For a comprehensive search across both, do both — they cover different corpora.3536## Examples3738```json39{"thought": "find recent constitutional AI papers", "tool": "semantic-scholar", "query": "constitutional AI alignment"}40```4142```json43{"thought": "what does the transformer paper build on?", "tool": "semantic-scholar", "paper_id": "arXiv:1706.03762"}44```4546```json47{"thought": "load BERT-architecture papers, just the top three", "tool": "semantic-scholar", "query": "BERT bidirectional transformer pretraining", "limit": 3}48```4950## Notes5152- Free public API; no key required. Set `SEMANTIC_SCHOLAR_API_KEY` for53 higher rate limits — unauthenticated use throttles quickly (429), and54 a throttled call reports itself as unavailable rather than as55 "no results".56- Papers with neither an abstract nor a PDF are dropped from search results.57- A paper S2 has no record of has no references here; extract them from58 the PDF itself in that case.
bdambrosio