Docling vs Kreuzberg: choosing structure and dependencies for local Office document extraction
Evaluate Docling first for an existing Python and DoclingDocument workflow; test Kreuzberg first when integrating extraction into applications in several languages and handling many legacy Office files. Compare structured output, OCR backends, and runtime dependencies, then verify offline boundaries with a cold start after preparing the models.
When an organization brings Word files, presentations, spreadsheets, and scans from a shared drive into a local document-processing workflow, the first decisions concern the structure downstream systems need, the runtime dependencies existing applications can support, and whether processing can restart without internet access.
If the team already builds post-processing around Python and DoclingDocument, evaluate Docling first. If it wants extraction inside Java, TypeScript, or Rust applications and fewer external conversion dependencies for legacy Office files, Kreuzberg deserves an initial trial. Both require acceptance checks against the actual formats and OCR configuration; “runs locally” alone does not establish offline readiness. This comparison draws on official documentation, not measured rankings of accuracy or speed.
A file inventory is more useful than a format count
Suppose an administration and engineering team needs to process policy DOCX files, training PPTX decks, XLSX registers, legacy DOC/PPT files, and scanned PDFs in the same batch. This is a selection scenario, not an executed test. Put each type on the acceptance checklist, and especially avoid treating successful DOCX processing as proof that legacy DOC has passed too.
Docling’s format documentation lists DOCX, XLSX, and PPTX, along with legacy DOC, XLS, and PPT, but the latter require LibreOffice. By contrast, Kreuzberg’s format reference says that since 4.3, legacy DOC/PPT extraction is native and no longer depends on LibreOffice. If old files arrive regularly, this directly changes the components a container must include. It does not establish which tool reconstructs every legacy file more completely.
Decide what to retain before comparing output
DoclingDocument combines text, tables, pictures, and section groupings in a unified structure, distinguishes the body from headers and footers, and carries location and provenance information when available. For teams preserving section relationships before generating text for different uses, it provides an explicit intermediate data model. The trade-off is that post-processing must understand that model rather than merely save one Markdown export.
Kreuzberg should not be reduced to “plain-text extraction” either. Its type reference defines DocumentStructure with parent-child relationships; nodes can represent headings, paragraphs, tables, and other content, while page numbers and bounding boxes may be null. Ask both tools to produce the structure you actually need, then check whether headings in training materials remain linked to the right content and register headers still correspond to their cells. Having structural fields differs from populating all of them for every file.
A more useful selection question is which structure downstream systems already consume and how much adaptation the other would require. If the only task is periodic text extraction for human search, a complex structure-conversion layer may not be worthwhile. If the workflow also archives by section and retains tables, structured results should be kept for the long term.
OCR choices change the deployment inventory
Docling’s OCR documentation lists engines including RapidOCR, EasyOCR, and Tesseract, with different language configurations. Choosing Tesseract requires traineddata for the languages actually used; installing the main package does not make every recognition language ready.
Kreuzberg’s OCR guide also offers several backends: Tesseract needs a system installation and language data, PaddleOCR can use a native backend, and EasyOCR is limited to Python. A Java business application therefore cannot assume that every OCR choice in a Python example transfers unchanged. For hybrid PDFs, also check separately whether pages with and without a text layer trigger the expected OCR behavior.
Runtime language deserves a separate evaluation. The Docling project provides Python usage and a command-line entry point. Kreuzberg’s installation documentation provides bindings for multiple languages, but notes that ONNX features on x86_64 require AVX/AVX2, and that the Windows Go, Elixir, and C/C++ bindings have a dynamic ONNX Runtime dependency. Native bindings can remove a service-integration layer, but they do not eliminate native libraries or hardware requirements.
Test offline operation with a full restart
Docling’s advanced options explain that models download on first use by default. They can be prepared with docling-tools models download and a local artifacts path. Remote model services require explicitly enabling enable_remote_services, but that switch controls sending user data, not model downloads. Third-party plugins have a separate allow_external_plugins switch. Check model assets, remote calls, and plugins individually rather than treating one switch as network isolation.
Kreuzberg’s CLI documentation provides kreuzberg cache warm to pre-download PaddleOCR and layout-detection models; prepare the corresponding assets when enabling other capabilities too. Its LLM integration can connect to either local inference engines or external providers, and VLM OCR sends page images to the selected model. A self-hosted program connected to an external endpoint is still not an entirely offline workflow.
Use the same acceptance plan for both tools: pin package versions, OCR backends, and model files, disconnect external networking in a fresh runtime with dependencies prepared, then cold-start processing for each sample type. Record missing resources, failed files, absent structure fields, peak memory, and completion time separately. Do not present cached extraction results as successful re-parsing. This is a proposed verification method, not a result measured for this article.
Ultimately, a Python workflow already using DoclingDocument should generally first check whether it can continue directly. Teams focused on integration inside multilingual applications, with many legacy Office files, can test Kreuzberg first. If nobody owns model assets, runtime-library updates, and offline regression checks, neither tool becomes maintenance-free simply by being installed locally.
Content note:This article is based on publicly available information and is intended as a reference for evaluating tools. Features, pricing, and related policies may change at any time; please refer to the latest official information.
Original source:docling-project.github.io