LoginSign Up
← Back to all sparks
D

Docling

AI-ASSISTANTS
Velocity7.5

Document parsing and conversion toolkit for AI pipelines

docling-project.github.io/docling ↗

Docling keeps widening what counts as a document — now video, charts, and agent skills.

document-conversionmultimodalchart-extractionagent-nativeocr
◆Current state
Docling ships roughly weekly, and each release adds input surface rather than polish. In six versions it picked up a video pipeline and InputFormat.VIDEO, legacy binary Office formats, an EBCDIC backend, a BoxNote backend, and native chart parsing across Word, Excel and PowerPoint that keeps the underlying data instead of a rendered image. Fix lists are long and concentrated in the DOCX, ODF and PDF backends, which is where format edge cases actually live.
◆Where it's heading
Two arcs run in parallel. The conversion core is becoming format-omnivorous — charts, video, mainframe encodings, archive formats — while the service layer grows the plumbing to run it at scale: chunking options and targets, generic batch connector sources, GCS, Azure Blob and Google Drive as both source and target. The agent skills in v2.118.0 point at a third arc: making Docling something an agent drives directly rather than a library a developer wires up.
◆Prediction
Expect the video pipeline to fill out using the ASR presets already in the tree, and the service layer to keep absorbing storage backends. The agent-skills entry suggests more agent-facing packaging is next.

◆Recent moves

  1. 3d ago

    v2.118.0 ships agent skills and an EBCDIC backend

    ⚡ SPARK

    Alongside an EBCDIC backend and RapidOCR language resolution across PP-OCR backbones, this release adds agent skills for using Docling — packaging the library's own usage knowledge for agents instead of leaving it to documentation. It extends the service-layer arc: Docling is being made callable by things other than a Python developer.

    View source ↗
  2. 7d ago

    v2.117.0 exposes chunking options and VLM logprobs

    Chunking options and targets move into the service datamodels, and VLM logprobs are exposed as generated tokens. Both serve callers building on top of Docling rather than end users — the service layer maturing in step with the conversion core.

    View source ↗
  3. 8d ago

    v2.116.0 makes the OCR pipeline layout-driven

    OCR becomes layout-driven with configurable modes, replacing one fixed approach with something tunable per document class. ASR also stops forcing English and auto-detects language instead — a small correction with outsized effect on non-English corpora.

    View source ↗
  4. 14d ago

    v2.115.0 adds batch connector sources and DOCX code-block detection

    The service client gains generic batch connector sources, continuing the at-scale plumbing arc, while the DOCX backend learns to detect code blocks from paragraph styles and fonts. Structural fidelity and operational reach advancing in the same release.

    View source ↗
  5. 17d ago

    v2.114.0 adds a video pipeline and native Word chart parsing

    ⚡ SPARK

    Docling crosses out of documents: VideoPipeline and InputFormat.VIDEO arrive with English-only and Distil-Whisper ASR presets, and Word charts join Excel and PowerPoint in being parsed with their data intact. Legacy binary Office formats land in the same release.

    View source ↗
  6. 23d ago

    v2.113.0 parses PowerPoint charts and adds cloud storage targets

    PowerPoint charts join the native-chart work, and the service exposes GCS, Azure Blob and Google Drive as both source and target types. The storage additions are what make Docling usable as a batch service rather than a local function call.

    View source ↗