Skip OCR when the PDF already has text

OCR is the expensive part of any PDF pipeline, and most PDFs do not need it. PdfToDocument reads the embedded text layer with word-level boxes, in the same pixel space PdfToImage renders at.

One pass to find out

import { Pipeline } from '@stabrise/scaledp'
import { PdfToDocument, hasUsableTextLayer } from '@stabrise/scaledp/pdf'

const probe = new Pipeline([new PdfToDocument({ resolution: 300 })])
const pages = await probe.transform(file)

const withText = pages.filter((row) => hasUsableTextLayer(row.document))
const needsOcr = pages.filter((row) => !hasUsableTextLayer(row.document))

hasUsableTextLayer(document, minimumBoxes = 1) is the whole decision. Raise minimumBoxes if your corpus contains PDFs whose "text layer" is a single stamped watermark.

Then OCR only what is left

import { PdfToImage } from '@stabrise/scaledp/pdf'
import { PaddleTextRecognizer } from '@stabrise/scaledp/ocr'

const ocr = new Pipeline([
    new PdfToImage({ resolution: 300, pageLimit: 0 }),
    new PaddleTextRecognizer({ keepFormatting: true }),
])

const scanned = needsOcr.length > 0 ? await ocr.transform(file) : []

Because both stages use the same resolution, the boxes from the text layer and the boxes from OCR are in one coordinate space — so the two sets of pages can be merged, drawn together, or fed to the same NER stage without conversion.

One pipeline, both columns

If you would rather not branch, run both readers and decide per row afterwards:

const pipeline = new Pipeline([
    new PdfToImage({ resolution: 300, keepInputData: true }),
    new PdfToDocument({ resolution: 300, outputCol: 'pdf_text' }),
])

Note that keepInputData: true on the first stage is required — PdfToImage deletes content by default, and PdfToDocument would then find nothing to read. And be aware this multiplies rows: two expanding stages over the same source column turn a five-page file into twenty-five rows. See Columns are the wiring.

For most corpora the two-pass version above is both cheaper and clearer.