DocumentDelta

Import

import { DocumentDelta } from '@stabrise/scaledp'

Group

Transform

Reads

document, document

Writes

document

Needs

nothing

A PDF page is often read twice: its text layer first, so results come in at once, then again with the text found inside its embedded images merged in by PdfMergeImageText. Running NER on the merged page a second time re-analyses every word the first pass already read. With a cloud NER service, it also pays for them twice.

With strategy: 'text-layer-wins' the merge only ever adds boxes, so the difference between the two readings is exactly the new text. DocumentDelta writes that difference as a Document:

import { DocumentDelta } from '@stabrise/scaledp'
import { PdfMergeImageText } from '@stabrise/scaledp/pdf'

new PdfMergeImageText({ outputCol: 'merged' }),          // keep `document` as the text layer
new DocumentDelta({ inputCols: ['document', 'merged'], outputCol: 'delta' }),

When the later reading added nothing (the page had no readable picture, or the text layer covered everything in it), the delta is an empty document. That is the common case on a born-digital PDF. Gate the NER that reads it with when: { path: 'delta.bboxes.length', gt: 0 } and that page costs nothing.

Merge the two NER readings back with NerMerger.

Boxes are matched by their text and rounded position, so a box that made a round trip through a service as JSON is still recognised. A first reading that failed counts as having read nothing, and the whole merged page is new.

Parameters