Page routing

Sending every page to a vision model reads anything and costs the most. Reading the text layer costs the least and returns nothing for a scan. Most folders are a mix, and complydoc routing works out which pages need what:

complydoc routing ~/contracts --no-ocr

Route   Pages  Documents
text       13          9
ocr         1          1
vision      5          5

Every page                             Cost
by the route it needs               $0.0121
from its text layer                 $0.0043
text layer, scans through local OCR $0.0067
as an image to a vision model       $0.0398

It writes complydoc-routing.json, a manifest an ingestion job can read, and the usual HTML report.

How each page is decided

Route	When
text	A usable text layer, and nothing on the page that plain text loses
ocr	No usable text layer, and a scan OCR can read
vision	Plain text would lose the page

A page takes the vision route when it carries a table with merged or stacked header cells, when it is mostly picture with a caption for a text layer, when it is a scan too coarse for OCR, or when OCR read it poorly. Every page carries the reason for its route, so a plan can be argued with.

The thresholds are in readiness.yaml under routing, and match the readiness signals they come from:

routing:
  min_characters: 40
  min_text_coverage_pct: 30
  picture_share_pct: 50
  min_ocr_dpi: 200
  min_ocr_confidence: 75
  vision_for_complex_tables: true

The manifest

{
  "model": "Claude Sonnet 5",
  "resolution": "medium",
  "counts": { "pages": { "text": 13, "ocr": 1, "vision": 5 } },
  "cost_usd": { "routed": 0.0121, "text_layer": 0.0043, "text_ocr": 0.0067, "vision": 0.0398 },
  "documents": [
    {
      "document": "vendor-assessment.pdf",
      "route": "vision",
      "pages": [
        { "page": 1, "route": "text", "reason": "a text layer covering 41% of the page", "characters": 1875 },
        { "page": 2, "route": "vision", "reason": "a table with merged or stacked header cells, which plain text loses", "characters": 640 }
      ]
    }
  ]
}

--print-json writes the same manifest to stdout and nothing else.

From Python

import complydoc as cd

report = cd.full_audit("./documents")
for document in report.documents:
    for page in document.routing.pages:
        print(document.relative_path, page.number, page.route, page.reason)

print(report.routing.routed_usd, report.routing.vision_usd)

cd.full_audit and complydoc audit compute routes too; the routing command exists to write the manifest and to price the mix on its own.

What the prices mean

The four figures use the model the report quotes and the folder's own pages. Pages routed to the text layer or to OCR are priced on the document's text tokens, shared between its pages by how many characters each holds, because a run does not count the tokens of a page separately. Pages routed to vision are priced as rendered images at the report's resolution.

A document that opens with no pages, such as an encrypted PDF, takes no route, and the summary counts those separately rather than leaving them out.