SmolDocling has layout problems on bilingual PDF

[mieslep]

Bug

META-NET PDFs (example Bulgarian] correctly finds text in the document but fails to detect title/section header, fails to understand the bi-lingual reading order (where columns span pages), and fails to pick up on a paragraph continuation to the next page.

The PDFs are authored in LaTeX (example Bulgarian).

Note there are 30 languages available, all with same format though different LaTeX approaches. https://european-language-equality.eu/meta-net-white-paper-series/

Steps to reproduce

import platform
from docling.datamodel.pipeline_options import (  # Additional imports for VLM
    VlmPipelineOptions,
    smoldocling_vlm_conversion_options,
    smoldocling_vlm_mlx_conversion_options,
)
from docling.pipeline.vlm_pipeline import VlmPipeline

if (
    "darwin" in platform.system().lower()
    and "arm64" in platform.machine().lower()
):  # optimized for Apple Silicon (MLX)
    vlm_options = smoldocling_vlm_mlx_conversion_options
else:
    vlm_options = smoldocling_vlm_conversion_options

vlm_pipeline_options = VlmPipelineOptions(
    force_backend_text=False,
    vlm_options=vlm_options,
)

converter_vlm = DocumentConverter(
    format_options={
        InputFormat.PDF: PdfFormatOption(
            pipeline_options=vlm_pipeline_options,
            pipeline_cls=VlmPipeline,
        )
    }
)

vlm_result = converter_vlm.convert("https://european-language-equality.eu/wp-content/uploads/2024/12/bulgarian.pdf")
vlm_doc = vlm_result.document

from docling_core.transforms.visualizer.layout_visualizer import LayoutVisualizer

layout_visualizer = LayoutVisualizer()
page_images = layout_visualizer.get_visualization(doc=vlm_doc)

display(page_images[1]) # title is two languages, split down the middle
display(page_images[4]) # two-column bilingual layout 
display(page_images[5]) # two-column bilingual layout continues
display(page_images[8]) # very long paragraph continues to next page
display(page_images[9])

Docling version

phil@Phils-MacBook-Pro docling % uv run docling --version
Docling version: 2.43.0
Docling Core version: 2.43.1
Docling IBM Models version: 3.9.0
Docling Parse version: 4.1.0
Python: cpython-312 (3.12.10)
Platform: macOS-15.5-arm64-arm-64bit

Python version

phil@Phils-MacBook-Pro docling % uv run python --version
Python 3.12.10

[dosubot]

SmolDocling currently doesn't have built-in support for detecting bilingual reading order, multi-column layouts, or paragraph continuation across pages in complex LaTeX-authored PDFs like the META-NET series. The pipeline and layout options don't expose configuration flags or heuristics for these cases, and the default layout model (DOCLING_LAYOUT_V2) doesn't address them either. The chunking and layout logic is modular, so it's possible to implement custom chunkers or reading order modules, but these aren't provided out of the box. There are also no documented workarounds or test coverage for these layouts in the codebase, so handling these documents would require custom development or future feature enhancements. For debugging, you can visualize cell bounding boxes and inspect the output of SegmentedPdfPage, but this won't resolve the underlying reading order or paragraph continuation issues.
Reference
Reference

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}