LoginSign Up
← Back to all sparks
D

Docling

AI-ASSISTANTS
Velocity5.0

Document parsing and conversion toolkit for AI pipelines

docling-project.github.io/docling ↗

Docling ships a native PDF pipeline and Apple Pages support while maturing its VLM integration layer.

document-parsingocrvlmpdfmulti-formatopen-source
◆Current state
Docling is releasing almost daily, expanding its supported document format surface with each version. Recent additions include a native PDF pipeline, completed Apple Pages support, JATS external hyperlinks, LaTeX output, and CUDA support for RapidOCR on ONNX Runtime. Threaded docling-parse is now the default, and VLM backends (Chandra, Unlimited-OCR) are getting tighter version management and error handling.
◆Where it's heading
Docling is converging on a universal document parsing pipeline — every major format (PDF, DOCX, PPTX, Apple Pages, AsciiDoc, WebVTT, EBCDIC, JATS) processed through a single interface, with multiple output formats (Markdown, JSON, LaTeX). The VLM integration is moving from experimental to production-ready as the team adds CUDA hardware support, min-version enforcement, and format-specific model handling. Scientific document workflows are an emerging target given the LaTeX output and JATS support.
◆Prediction
The experimental two-stage VLM pipeline will stabilize into a production path in the next few releases. Scientific document ingestion (LaTeX, JATS, math-heavy PDFs) will likely become a promoted use case as the output format support deepens.

◆Recent moves

  1. 11d ago

    v2.126.0: native PDF pipeline, Apple Pages complete, JATS hyperlinks

    v2.126.0 adds a native PDF pipeline (a new processing path distinct from the existing backends), completes Apple Pages document parsing, preserves external hyperlinks in JATS documents, and fixes ordered list numbering in Markdown output. Broadens both input and output fidelity.

    View source ↗
  2. 12d ago

    v2.125.0: CUDA RapidOCR, VLM min-version check, HTML table fixes

    v2.125.0 enables CUDA for RapidOCR on ONNX Runtime, adds min-version enforcement for VLM engines (fixing an MLX compatibility issue), improves HTML table parsing inside figure elements, and fixes WebVTT multi-line cue ordering. Hardware acceleration and format correctness both advance.

    View source ↗
  3. 15d ago

    v2.124.0: LaTeX output format added

    v2.124.0 adds LaTeX as an output format from the CLI — a meaningful new target for scientific document workflows where Markdown isn't rich enough. Terminal task callback events and Retry-After header handling for the service client round out the release.

    View source ↗
  4. 18d ago

    v2.123.1

    v2.123.1 patches VLM spacing for Chandra br tags, defers heavy CLI imports for lightweight installs, and drops the UTF-8 BOM from CSV first cells. Bug fixes; no new capabilities.

    View source ↗
  5. 20d ago

    v2.123.0: threaded parsing now default across all interfaces

    v2.123.0 makes threaded docling-parse the default across SDK, CLI, service, and extraction — a performance improvement that now applies everywhere without explicit opt-in. Layout improvements for nesting tables and pictures in form regions also ship.

    View source ↗
  6. 22d ago

    v2.122.0: compact_tables, Apple Pages heading recovery, layout speed

    v2.122.0 adds a compact_tables option, recovers Apple Pages titles and headings (completing the iWork story that finalizes in v2.126.0), and speeds up layout and table cell matching. A broad-front quality improvement.

    View source ↗