Guide

How it works

Page Search reads every page of a document and judges, for each of your questions, whether the page helps answer it.

Every page is read

There is no index and no ranking that decides which pages are looked at. A model reads each page on its own and gives each question a relevance: the chance, from 0 to 1, that the page helps answer it. All pages are read at the same time, so a search takes about as long as one page: around a second for a 160-page report.

Most pages are read as text, with their tables, columns and headings kept. A page whose information is drawn rather than written (a chart, a diagram, a scanned page, a form) is read from its image as well. The response of an upload tells you how many pages are read that way (image_pages). It does not change the price.

Questions

A search takes 1 to 10 questions. Each page is judged against all of them in one go, so ten questions take as long as one, and cost less than ten searches. Ask the questions as you would ask a person: "What was capital expenditure in 2018?", "Which pages describe the PFAS litigation?".

Relevance

A page's relevance is the chance it helps answer the question: 0.95 means it almost certainly does, 0.05 that it almost certainly does not.

• By default each question lists the pages likely to help, most relevant first: as many as the question needs, and always at least 3 (every page, for a shorter document). An answer spread over several sections gets more pages.
• To get a set number of pages, pass top_k: the most relevant pages, exactly that many.
• When every page listed has a low relevance (say, under 0.1), the document probably does not answer the question.

The answer to a question is often on more than one page, such as a figure in a summary table and again in a note.

Documents

• Upload once, search many times. A document is processed when it is uploaded; searches reuse it.
• The same file is one document. Uploading a file you already have returns that document, with "duplicate": true, and does not process it again.
• Unused documents expire. A document expires 30 days after its last search (expires_at), and each search keeps it another 30 days. Searching an expired document returns document_expired: upload it again.
• Storage is free up to 100,000 pages at a time. Past that, uploads are refused until you delete documents or they expire.
• Deleting a document removes it and everything made from it.

Billing

A search costs its document's pages times the price for its number of questions (see pricing); pages that failed are not charged. Each response says what it cost (usage.cost_usd) and what credit is left (usage.balance_usd). When your credit runs out, uploads and searches return insufficient_balance.