Data Sources

Rivul AI does not write your citations. References come from the scholarly metadata systems below, formatted by a deterministic rule, so a citation can be traced to the index that produced it. This page lists every one of those systems, the licence Rivul AI reads it under, and what Rivul AI keeps. Nothing here is paid for and no scholarly source charges Rivul AI for access.

Used for similar-paper recommendations, and as a fallback literature search, with arXiv, when the Rivul Index is unavailable.

Read under the Semantic Scholar API licence, which requires attribution to Semantic Scholar and a link back to the paper. Every result that comes from Semantic Scholar carries that link. The underlying graph data is generally ODC-BY; individual records may carry their own terms, and open-access PDF links must be checked at the source.

Rivul AI stores title, authors, year, venue, abstract, DOI, arXiv id and citation count for papers you save to your library. Search responses are cached for 24 hours and recommendations for 7 days, then discarded.

Used for preprint search and for fetching open-access PDFs.

arXiv metadata is CC0 1.0; the e-prints themselves stay under their authors' own licences, so a PDF Rivul AI fetches for you is the author's work under the author's terms, not Rivul AI's. Thank you to arXiv for use of its open access interoperability.

arXiv asks for no more than one request every three seconds across all machines a service controls. Rivul AI holds itself to that rather than treating each server as a fresh allowance: requests are serialised and each one has to claim a three-second slot from shared state. Query results are cached for a day, which arXiv's manual asks integrators to do.

Rivul AI stores the metadata, the extracted text of any PDF you fetch (which is what grounds chat and citation checking), and the PDF file itself in Rivul AI's storage so your library keeps working offline and reopening a paper does not re-download it. Deleting the library entry deletes both.

Used for DOI lookup, publisher metadata, and retraction and preprint status.

Crossref asserts no licence over the metadata facts it distributes. Its API asks callers to identify themselves with a contact address, which puts them in the polite pool with a published concurrency limit; Rivul AI identifies itself and stays inside that limit.

Rivul AI stores the reference fields a citation needs — DOI, title, authors, year, venue, volume, issue, pages, publisher, type — and a retraction flag on library entries.

Used to flag retracted papers in your library.

Crossref acquired the Retraction Watch database and publishes it as a CC0 dataset, rebuilt every working day. Rivul AI reads the whole dataset once a night and compares it against the DOIs in its users' libraries, rather than polling for one DOI at a time, so every paper in a library is checked every night regardless of when it was added.

Rivul AI stores only the resulting flag. The dataset itself is streamed, matched and discarded; no copy is kept.

The Rivul Index, the primary search behind Search papers, Research, and Rivul MCP, is built from a dated OpenAlex snapshot of more than 300 million scholarly works. It supplies titles, authors, years, venues, abstracts, citation counts, DOIs, and source links.

OpenAlex data is CC0, which requires no attribution; it is credited here because citing a source is the practice this product is for. Records are only as current and correct as OpenAlex was on the snapshot date, so check a record against its source before relying on it. When the index is unavailable, search falls back to Semantic Scholar and arXiv and the interface says results may be partial.

Rivul AI keeps the snapshot in its own search service and stores citation counts and abstracts on saved library entries.

Used as a fallback for DOIs that Crossref does not hold — arXiv DOIs, datasets and software are registered with DataCite.

DataCite metadata is CC0.

Rivul AI stores the same reference fields as for Crossref.

Used to find a legal open-access copy of a paper by DOI.

Unpaywall's data is CC0. Its API requires a real contact email, and Rivul AI sends one; when none is configured Rivul AI skips the lookup rather than presenting an address that does not receive mail.

Rivul AI stores nothing from Unpaywall beyond a cached answer to "is there an open-access copy of this DOI", kept for a week for a hit and a day for a miss.

Your drafts, your notes, your annotations and your account details are never sent to any of the sources above. What leaves Rivul AI for them is a search query, a DOI, or an arXiv identifier.

If you maintain one of these services and something on this page is wrong or out of date, tell us through the contact form at rivul.ai/contact and it will be corrected.