Text is lost when a `TJ` array is split across two content streams of a page

Aperta
  • Qualcosa non va?
#450 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 5 giorni

@morten-lagabote ci sta già lavorando.

Dal 6/11/2026.

  • #451 di @morten-lagabote — aperta

Valutazione

Difficoltà
4/5
Tempo stimato
3-5 giorni
Idoneità per principianti
25/100
Tipo di issue
Bug
Chiarezza
Specificata chiaramente
Stato di attività
Ferma
Stack tecnologico
cpp
Ambito
backend

Direzione di ricerca

Start with pdf_decoder< PAGE >::decode_contents in src/parse/pdf_decoders/page.h, identified as the source of the issue, and trace how it tokenizes each stream in a page’s /Contents array. Reproduce with the provided Python example, especially the split after [. Done when text in a TJ array spanning two streams is preserved; an open linked pull request (#451) is already addressing the issue.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

Summary

When the /Contents of a page is an array of streams, one stream can end inside a token: for example after the [ of a TJ array, with the strings and ] TJ at the start of the next stream. ISO 32000-1 (7.8.2) allows this, because the streams are one content stream that is divided at token boundaries. docling-parse loses the text of such a TJ. pypdfium2 keeps it.

A split between two operators, or between the array and TJ, is read correctly.

The cause is in pdf_decoder<PAGE>::decode_contents (src/parse/pdf_decoders/page.h): it tokenizes each stream of the array on its own, so an array that is open at the end of a stream is not completed by the next stream.

I can open a PR.

Reproduction

Self-contained, docling-parse 7.24.2 (Python 3.12, macOS arm64):

import pypdfium2
from docling_parse.pdf_parser import ContentConfig, ContentLevel, DoclingPdfParser


def make_pdf(path, first, second):
    # One page whose /Contents is an array of two streams.
    objs = [b"<< /Type /Catalog /Pages 2 0 R >>", b"<< /Type /Pages /Kids [3 0 R] /Count 1 >>",
            b"<< /Type /Page /Parent 2 0 R /MediaBox [0 0 595 842] /Contents [4 0 R 5 0 R] "
            b"/Resources << /Font << /F1 6 0 R >> >> >>",
            b"<< /Length %d >>\nstream\n" % len(first) + first + b"\nendstream",
            b"<< /Length %d >>\nstream\n" % len(second) + second + b"\nendstream",
            b"<< /Type /Font /Subtype /Type1 /BaseFont /Helvetica >>"]
    out, offsets = bytearray(b"%PDF-1.5\n"), []
    for i, o in enumerate(objs, 1):
        offsets.append(len(out))
        out += b"%d 0 obj\n" % i + o + b"\nendobj\n"
    xref = len(out)
    out += b"xref\n0 %d\n0000000000 65535 f \n" % (len(objs) + 1)
    for o in offsets:
        out += b"%010d 00000 n \n" % o
    out += b"trailer\n<< /Size %d /Root 1 0 R >>\nstartxref\n%d\n%%%%EOF\n" % (len(objs) + 1, xref)
    open(path, "wb").write(out)


head = b"BT /F1 12 Tf 72 720 Td [(Line one)] TJ 0 -20 Td "
tail = b" 0 -20 Td [(Line three)] TJ ET"
for name, first, second in [
    ('split after "["', head + b"[", b"(Line two)] TJ" + tail),
    ('split before "TJ"', head + b"[(Line two)]", b"TJ" + tail),
    ("split between operators", head + b"[(Line two)] TJ", tail),
]:
    make_pdf("/tmp/split.pdf", first, second)
    page = DoclingPdfParser(loglevel="fatal").load("/tmp/split.pdf").get_page(
        1, content_config=ContentConfig(line_cells_content_level=ContentLevel.COMPUTE_AND_MATERIALIZE)
    )
    pdfium = pypdfium2.PdfDocument("/tmp/split.pdf")[0].get_textpage().get_text_range().splitlines()
    print(f"{name:24} docling-parse: {[c.text for c in page.textline_cells]}  pypdfium2: {pdfium}")

Output:

split after "["          docling-parse: ['Line one', 'Line three']  pypdfium2: ['Line one', 'Line two', 'Line three']
split before "TJ"        docling-parse: ['Line one', 'Line two', 'Line three']  pypdfium2: ['Line one', 'Line two', 'Line three']
split between operators  docling-parse: ['Line one', 'Line two', 'Line three']  pypdfium2: ['Line one', 'Line two', 'Line three']
Where we see it

In the regression dataset, 11169157880668894066-1.pdf, page 1: content stream 1 ends with [, and the word fever(A:1), is lost.

In decisions of the Supreme Court of Norway, published as PDF on domstol.no. In each, a content stream of page 1 ends with [, and the lost text is a judge's name or a part of the case line:

Document Text lost on page 1
HR-2019-1786-A dommer Wilhelm Matheson
HR-2026-1454-S dommer Bergljot Webster
HR-2026-780-A Anke over Borgarting lagmannsretts dom 11. april 2025
HR-2023-753-U Natt til søndag 6. februar 2022 ca. kl. 02.30 utenfo (the line starts with r Bunnpris Torggaten i)
HR-2025-1883-U U, (sak nr. 25 (the line is HR-2025-1883- and -137197STR-HRET), straffesak, anke over beslutning:)

These five PDFs are made by Microsoft Word 2016, and page 1 of each has 8 content streams.

Lingua principale
C++
Stelle
346
Fork
92
Merge medio
4g 3h
PR unite (30g)
26

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di docling-project/docling-parse

Tutte le issue di docling-project/docling-parse

Issue simili

Altre issue su C++

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.