Skip to main content
Paper search finds papers by title and abstract. Many papers have no abstract in the index, so you read the paper itself. You search, keep the papers with an open PDF, and read the first pages of each with Scrape. You need an API key in STOPHY_API_KEY. See the Quickstart. Paper search needs an API key.

1. Search for papers

This request finds papers from 2024 on, newest first, and asks for 3.
results holds the papers. This is the real answer. The abstracts are cut, and two of the three papers had none.
primaryId is the paper’s best outside id, and openAccessUrl is where a free copy lives. It is a PDF for the first and third paper, and a landing page for the second. See Papers for every option.

2. Read a paper

Scrape reads a PDF the way it reads a page and returns Markdown. This request reads the first 2 pages of the third paper and cuts the text at 2,000 characters.
This real answer is cut. The middle of the text is left out.
totalPages is the pages in the file, 9 here. nextOffset says where to read on. Send the same request with "offset": 1969. Send a larger maxPages for more pages, and pageMarkers: true to see where each page starts. See PDFs and documents. The read cost $0.0015. A landing page such as the second paper’s is not a PDF. Scrape reads it as a page, and it may not hold the paper. Keep the papers whose link ends in .pdf or points at arxiv.org/pdf.

3. Read every paper

This function searches, keeps the papers with a direct PDF link, reads the first 2 pages of each, and returns notes with the id, title, year and text. The id is the paper’s primaryId, so a citation points at the same paper wherever you saw it.
Each note is ready to hand to a model with the instruction to quote only from text and cite id. For a paper you hold as a file, Parse documents gives the same Markdown from an upload.

What to check

  • results is not empty, and the papers match your topic. If not, change the query, or add authors, from and to.
  • Every note has a text that starts with the paper’s title. If a PDF comes back empty, try "mode": "ocr" in the PDF options.
  • The text you hand a model is cut at maxChars. Read on with nextOffset when the answer needs the methods or the results.
  • A claim in your summary is a quote you can find in text. Check three.

What it costs

A page that cannot be read costs nothing. See Pricing.