base DocumentConverter misses all text

[mieslep]

Bug

META-NET PDFs (example Bulgarian] find no text in the document.

The PDFs are authored in LaTeX (example Bulgarian).

Steps to reproduce

# Create a converter instance
converter = DocumentConverter()

# Convert a document
result = converter.convert("https://european-language-equality.eu/wp-content/uploads/2024/12/bulgarian.pdf")
doc = result.document

# statistics
print(f"Document title: {doc.name}")
print(f"Number of pages: {len(doc.pages)}")
print(f"Number of tables: {len(doc.tables)}")
print(f"Number of pictures: {len(doc.pictures)}")
print(f"Number of texts: {len(doc.texts)}")

gives:

Document title: META-NET
Number of pages: 89
Number of tables: 14
Number of pictures: 39
Number of texts: 0

But with SmolDocling VLM:

Document title: Document
Number of pages: 89
Number of tables: 13
Number of pictures: 17
Number of texts: 946

(though it takes a lot longer, 9.7 sec/page vs 0.67 sec/page on M4 MacBook Pro, and there are other detection problems that are separate from this issue).

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]

The base DocumentConverter misses all text in many LaTeX-authored PDFs because it relies on standard PDF text extraction, which only works if the PDF contains a proper text layer. LaTeX-generated PDFs often encode text as vector graphics or use non-standard encodings, so the extractor finds zero text objects. SmolDocling VLM works because it uses a vision-language model (VLM) and OCR to "see" and transcribe text from page images, which is much more robust for these cases but slower due to the extra processing involved.

You can improve extraction in the base DocumentConverter by enabling OCR. Here’s how to do it:

from docling.document_converter import DocumentConverter
from docling.datamodel.pipeline_options import PdfPipelineOptions
from docling.datamodel.input_format import InputFormat
from docling.datamodel.format_option import PdfFormatOption

converter = DocumentConverter(
    format_options={
        InputFormat.PDF: PdfFormatOption(
            pipeline_options=PdfPipelineOptions(
                do_ocr=True,
                force_full_page_ocr=True,  # Forces OCR on every page, even if a text layer is present
                # Optionally, specify the OCR engine and language:
                # ocr_options=TesseractCliOcrOptions(lang=["eng"])
            )
        )
    }
)
result = converter.convert("https://european-language-equality.eu/wp-content/uploads/2024/12/bulgarian.pdf")
doc = result.document
print(f"Number of texts: {len(doc.texts)}")

Supported OCR engines include EasyOCR (default), Tesseract, and RapidOCR. You may need to install and configure these separately, depending on your environment. Enabling full-page OCR will significantly slow down processing (similar to the VLM pipeline), but it should recover text from PDFs where the base extractor fails. For best results, make sure the OCR engine is installed and working, and set the language option if your document isn’t in English. More details and troubleshooting tips are available in the Docling OCR configuration discussion and OCR setup guide.

In summary: the base converter is fast but fragile for LaTeX PDFs; enabling OCR or using the VLM pipeline is slower but much more robust for extracting text from visually complex or non-standard PDFs.

To reply, just mention @dosu.

---

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