LoginSign Up
← Back to all sparks
D

Docling

AI-ASSISTANTS
Velocity6.3

Document parsing and conversion toolkit for AI pipelines

docling-project.github.io/docling ↗

Docling keeps swallowing new formats, and now the parsing engines behind them are swappable.

document-parsingformat-coveragepluggable-enginesocrstructure-inference
◆Current state
Docling releases every three to four days, alternating feature drops with tight fix releases. The current one is purely corrective: DOCX headings detected by outline level when the style is not literally named Heading, Markdown tables keeping their last cell without a trailing pipe, and the service client serializing engine options in full. Format coverage now spans PDF, Office, ODF, HTML, JATS, email, audio and video.
◆Where it's heading
The engine layer is where the interesting movement is. Docling is shifting from one opinionated pipeline to a set of interchangeable layout, table and OCR backends the caller picks per run, which turns the library into a harness for models rather than a fixed parser. A second thread: the project shipped agent skills for itself in v2.118.0 and a separate docling-client package in v2.120.0, both pointing at being consumed programmatically rather than only imported. The structural-inference work — heading levels from font weight, now from DOCX outline levels — shows the parser learning to read documents that never declared their own structure.
◆Prediction
Expect the engine-selection surface to keep widening, with OCR joining layout and table structure as a CLI-selectable backend. The steady stream of format-specific crash fixes suggests coverage is outrunning hardening, so more of these short corrective releases are likely between feature drops.

◆Recent moves

  1. 1d ago

    DOCX outline-level headings and Markdown table cell fixes

    Three corrective changes: DOCX headings are detected via outline level when the style carries a custom name, Markdown table rows keep their final cell without a trailing pipe, and the service client serializes engine options completely. The DOCX fix extends the same structure-inference thread as the font-weight heading work, but as a repair rather than a new capability.

    View source ↗
  2. 4d ago

    Release CI fix, no user-facing changes

    A single-line patch to the release CI, issued hours after v2.120.0. Nothing reaches the parsing pipeline.

    View source ↗
  3. 4d ago

    Heading levels inferred from font weight; pluggable CLI engines

    Two threads land together: PDF heading levels inferred from typography rather than markup, and --layout-engine plus --table-structure-engine exposing the pluggable backends at the command line. The separate docling-client package continues the split between the library and the way callers reach a running service.

    View source ↗
  4. 8d ago

    Outlook .msg support and Unlimited-OCR grounding

    Outlook .msg joins the format list with optional attachment listing, extending the email backend into the proprietary Microsoft container. Parsing Unlimited-OCR grounding output continues the pattern of adding OCR engines rather than committing to one.

    View source ↗
  5. 11d ago

    Layout label and PDF picture-in-table fixes

    A fix-only release: hyphenated layout labels normalized, pictures inside table cells preserved, and the HTML renderer stopped from making implicit file requests. The last of those is a quiet hardening change worth noting for anyone rendering untrusted documents.

    View source ↗
  6. 15d ago

    EBCDIC backend, docling agent skills, all PP-OCR languages

    An EBCDIC backend pushes coverage into mainframe territory, and the RapidOCR model refactor resolves all PP-OCR languages by version and backbone rather than enumerating them. Shipping agent skills for using Docling is the first sign the project treats agents as a first-class caller.

    View source ↗