Semantic Scholar

Everything is free. The FAQ calls Semantic Scholar a free, AI-powered search and discovery tool and says no account is needed to reach the papers. A free account adds a library with folders, bulk citation export and shareable folder links, email alerts for authors, papers, topics and Research Feeds, the feeds themselves, and the option to claim an author page.

Signing in with an institution through OpenAthens, eduGAIN or InCommon adds an Access PDF via Institution button where the publisher takes part in GetFTR, or a LibKey route where it does not. Without that, a paywalled paper links out to the publisher, where reading it may need a subscription or a purchase.

Machine learning ranks results, assigns fields of study, flags highly influential citations and classifies why one paper cites another. TLDRs, the one-sentence AI summaries shown beside results, are limited to computer science and biomedical papers. The generative features sit mostly in Semantic Reader, which opens only arXiv papers, 250k+ of them so far:

• Ask This Paper sends the paper and your question to OpenAI’s gpt3.5-turbo-16k and shows supporting snippets from the paper; it has been tested only on English-language papers.
• Clicking an underlined term asks ChatGPT for a definition in context.
• Skimming highlights mark Goal, Method and Result passages on a limited set of English papers, on desktop.
• Topic pages use large language models to extract topics and write their definitions.

The FAQ warns that generated text can contain factual errors that are hard to spot, from subtle to serious, so check each answer against the snippet it cites.

The API page counts 214 million papers, 2.49 billion citations and 79 million authors, while the homepage search bar showed 239,525,222 papers; the FAQ says the corpus keeps growing and points to the search bar for the current number. Content comes from web indexing and from partnerships with publishers and indexes such as PubMed, arXiv and Springer Nature.

The gaps are documented: book coverage is very limited, patents are not included and the focus is on English-language work, so citation counts and h-index values will differ from other sites. The FAQ itself advises against using its h-index for comparative research assessment. Other research tools build on this corpus; the Undermind profile covers a deep-search service that runs on the Semantic Scholar database.

Search does not support Boolean operators or wildcards, though quoted text works, and abbreviations such as KDD are not expanded, so search with full terms. Results can be sorted by relevance, citation count, most influential papers or recency, and the reranker model is published on GitHub.

A Research Feed is tied to a library folder and uses a paper embedding model. Papers saved in the folder count as positive signals and papers marked not relevant in the feed as negative ones. The feed refreshes daily and recommends only papers published in the last 3 months, so it is a monitoring tool rather than a way to find older literature.

Most endpoints work without a key, but those calls share 1,000 requests per second among all unauthenticated users and can be throttled at busy times. The API tutorial says an individual key gets 1 request per second across all endpoints and recommends batch and bulk endpoints for large jobs. The API documentation sets the working limits:

• Paper batch: 500 paper IDs per call, up to 10 MB per response.
• Relevance search: up to 1,000 ranked results for a query.
• Bulk search: 1,000 papers per call, up to 10,000,000 papers in total.
• Author batch: 1,000 author IDs per call.

The API license requires an attribution to Semantic Scholar wherever results are published and a citation in scientific papers that use the API, and says exceeding or trying to get around the rate limits can lead Ai2 to suspend or disable access.

1. List ten papers from your field, including two books or chapters and, if your field has them, two non-English papers.
2. Search each title in quotes and record which appear, with their citation counts next to the counts in another index you use.
3. Save five to a library folder, switch on a Research Feed and mark three recommendations as not relevant.
4. Over the next week, count how many daily recommendations you would actually read.
5. On one arXiv paper, ask Ask This Paper a question you can answer yourself and compare the answer with its supporting snippet.

Accept if your core papers appear and the generated answer matches its snippet. Reject if key papers or whole publication types are missing, or the answer states something the snippet does not. No result is recorded here.