Splitting multi-invoice PDFs by invoice-number label

Before extraction, the text of every page is read and tested for an invoice-number label followed by a value. The labels come from the same lexicon packs the extractor uses — Invoice No., Invoice Number, Invoice #, Bill No., Document No., and from the German pack RECHNUNG, Rechnungsnummer, Belegnummer, Beleg-Nr.. The value may be separated by a colon, dot, hash or dash and must be at least three characters of letters, digits, slashes or dashes. In code:

export function pageLooksLikeInvoiceStart(text: string, labels: readonly string[]): boolean {
  if (!text.trim()) return false
  for (const label of labels) {
    const re = new RegExp(`${escapeRegex(label)}\\s*[:.#\\-–]?\\s*([A-Za-z0-9][A-Za-z0-9\\/\\-]{2,})`, 'i')
    if (re.test(text)) return true
  }
  return false
}

A page that passes starts a new invoice. A page that fails belongs to the invoice before it. Pages before the first start — the cover sheet, a suite directory, a summary — are dropped. The split only happens when at least two start pages are found; a PDF with one invoice is ingested whole.

We generated a five-page PDF with one line of text per page and ran the splitter on it. The page texts were:

[ 1, 2, 4 ]
pack-01.pdf 1 page(s)
pack-02.pdf 2 page(s)
pack-03.pdf 1 page(s)

Three documents: the English invoice, the two-page German invoice (Seite 2 von 2 has no label, so it stays with the RECHNUNG page), and the Belegnummer invoice. The summary page produced nothing. Each part keeps the original pages intact — it is rebuilt with pdf-lib from the source pages, not re-rendered — so extraction sees exactly what the supplier printed.

• Titles are not labels. “TEST INVOICE 1” on a directory page is a heading, not a field; it has no value after a label, so it is ignored. That one distinction removes most false starts in packs that begin with an index.
• Continuation pages take care of themselves. A second page of line items has no invoice-number field, so it attaches to the invoice before it. Nothing has to detect “page 2 of 2”.
• Languages are a lexicon concern, not a splitter concern. Adding a Danish or French label is one line in a pack file; the splitter never changes.

• If the PDF has no readable text (a pure image scan), or the split throws for any reason, the original file is ingested unchanged — it is never dropped.
• Each part becomes its own document and counts toward the batch limit (40 files per upload by default). A pack that splits into more parts than that reports the extras as ignored rather than failing the upload.
• Files arriving through the SFTP inbound connector are not split; each file there is treated as one document.
• A page that carries two invoice-number labels (a credit note referencing the original invoice, for example) still starts exactly one document — the split is per page, never within a page.