Key points of a document

The reading bottleneck this replaces

Long documents create a specific kind of bottleneck: someone qualified has to read the whole thing before anyone can act, and that person's time is usually the most expensive part of the workflow. A procurement lead skimming a supplier agreement, a case manager reviewing a settlement letter, a compliance officer scanning a policy update, all of them are really looking for a handful of sentences buried in pages of boilerplate. This endpoint exists to surface exactly that handful, automatically, so the qualified reader spends their time deciding rather than searching.

What kind of points it pulls out

You send a document to POST /ocr/key-points and receive a task_id right away, since identifying what actually matters in a document takes real reading, not a keyword scan. The result groups findings into practical categories: decisions or commitments stated in the text, figures such as amounts, dates, and quantities that carry weight, and risks or obligations worth a second look, like penalty clauses, deadlines, or unusual terms. You choose how the result reaches you, a signed webhook fired the moment it is ready, or a signed link you fetch within 24 hours.

Why this is judgment, not just summarization

A generic summary compresses a document; key-point extraction is a different task, it filters for what is actionable rather than what is representative. A summary of a contract might mention that it covers payment terms; key points identify the actual net-30 clause and the late-fee percentage, because those are the numbers someone needs to negotiate or approve. That distinction matters most in dense, formal writing, contracts, policy documents, regulatory filings, where the important content is a small fraction of the total text and burying it is often the point of the drafting style.

How it slots into a review workflow

Key-point extraction fits naturally after a classification step, so contracts get read for clauses and risks while invoices get sent to structured extraction instead, and before a human approval step, where the extracted points become the summary someone actually reads before signing off. Because it is cheap per call and stateless, teams often run it on every incoming document as a first pass, reserving full manual review for the documents where the extracted risks warrant it.

Cost built for high-volume reading

Pricing is $0.003 per request plus $0.0135 per page, kept deliberately low because this is meant to run on volume, across every contract or policy a team receives, not just the occasional flagged document. A failed task is never charged, we retry automatically up to three times before returning a clear error explaining what went wrong. There is no free tier or trial, access requires prepaid balance and calls without one return HTTP 402, which keeps the endpoint responsive instead of congested by unpaid traffic.