Repository navigation
webvtt: text after a multi-line span is attached to the wrong cue line #4104
Description
Activity
Hi @businessarshgoyal! I'm Dosu and I'm helping the IBM Docling team.
I can confirm this bug in the current
maincode. In_extract_components,paris bound tocue_text[-1]once at the top of the call. When a span like<v Bob>Hello\nthere</v>contains a line terminator, the recursive call into_extract_components(component.internal_text.components)appends a newAnnotatedParto the sharedcue_textlist and starts writing into it. But when the recursive call returns, the outer loop'sparstill references the original paragraph. Any sibling component processed afterward (like the plain-text" and afterwards") therefore gets appended to the stale first paragraph instead of the current one — exactly the misordering you described.The existing tests don't cover this scenario:
test_multi_line_cue_text_preservedonly uses plain-text multi-line cues, andtest_cue_ids_plus_multiline_with_voice_and_stylehas voice/style spans but no line break inside a span's payload. So your repro is a legitimate gap.The fix should be straightforward — re-fetch
par = cue_text[-1]after the recursive_extract_componentscall returns (line 155), so the outer scope always writes into whatever paragraph is currently last incue_text. Something like:parents.append(item) _extract_components(component.internal_text.components) parents.pop() + par = cue_text[-1] # the recursive call may have started a new paragraph
I don't see an existing PR addressing this specific ordering bug yet.
To reply, just mention @dosu.
Share context across your team and agents. Try Dosu.
Bug
In
WebVTTDocumentBackend.convert,_extract_componentscaches the current paragraph in a local variable (par = cue_text[-1]) before iterating over a cue payload. When a nested span (voice, bold, italic, underline) wraps a line terminator, the recursive call starts a new paragraph incue_text, but the caller keeps writing into its stalepar. Everything that follows the span on the next line is therefore appended to the first line's paragraph, so the cue text comes out in the wrong order.Steps to reproduce
sample.vtt:Actual
export_to_markdown()rendersHello and afterwardsfirst andthereafter it.Expected
and afterwardssits on the same cue line asthere, so it belongs to that line's paragraph.Docling version
main (2.x), Python 3.12