Skip to content

LearnIngest & extract

Page routing

Per-page extraction routing

Choosing an extraction method separately for each page: keep the text layer, run OCR, or send the page to a vision model, based on what that page needs.

In DocSlurpGlossary entry

On this page

Page routing chooses how to extract each page of a document separately, instead of running one method on the whole file.

The options form a cost ladder:

Route When it fits Relative cost
Keep the text layer born-digital text with mapped glyphs almost free
OCR scanned page of plain text low, local CPU
Vision-language model tables, charts, math, handwriting, messy scans high, model call
Skip blank page free

A 40-page board-update-2026.pdf might be 36 pages of clean digital text, two scanned appendix pages, one page with a dense table and one blank separator. Sending all 40 to a vision model costs 40 model calls. Routing keeps the text layer for the 36, runs OCR on the two scans, sends the table page to a vision model and skips the blank one: one model call instead of 40. The output is as good or better, because the text layer for those 36 pages is exact.

A router needs cheap signals for every page before deciding: does it have a text layer, do its glyphs map to real characters, how much of the page is covered by images, does it contain math symbols or ruled table lines? The expensive reader is called only where those signals say the cheap one will fail.

The gotcha: a router’s mistakes are silent. A page wrongly sent down the cheap path produces plausible but wrong text, such as columns read straight across, and nothing downstream flags it. Run new routing rules in shadow mode first, logging what they would have chosen, and compare against a stronger reader on a sample before acting on them. See pages per $1.

In DocSlurp

In DocSlurp

By default classify-route authorizes OCR only for pages with no native text whose images hold the content, and fast never reads pages (core/worker/src/stages/classify-route.ts). An opt-in page router (pageRouter.mode: off, shadow or active) probes each page and plans roles such as accept-text, ocr, table and math (core/worker/src/routing/planner.ts).

Sources

Jump to