Docling adds video, XBRL, and email parsing as document AI eats every format

The Docling project shipped a batch of new parsers this cycle, and the list says more about where document AI is heading than any single feature does. The converter, started by the AI for knowledge team at IBM Research Zurich and now hosted under the LF AI & Data Foundation, already handled PDF, DOCX, PPTX, XLSX, HTML, LaTeX, and a long tail of office formats. The new additions are video files (MP4, AVI, MOV, MKV, WebM) with ASR transcripts and representative keyframes, ODF text, spreadsheets, and presentations, XBRL financial reports, EML and MSG email, EPUB, Apple Pages across both container generations, plain text and Markdown supersets like .qmd and .Rmd, plus chart understanding that turns bar, pie, and line charts into tables or code.

That is a wide net. The interesting part is not the count. It is the direction.

The format war is over, and the parser won

For the past two years, the bottleneck in retrieval-augmented generation has not been the model. It has been ingestion. Every enterprise RAG pilot hits the same wall: the documents that matter live in a dozen formats, and the ones that carry the most value (financial filings, scanned contracts, recorded meetings, email threads) are the hardest to parse cleanly. A model that reads Markdown perfectly is useless if the source is a 200-page XBRL filing or a WebM screen recording of a quarterly review.

Docling’s answer is to absorb the format problem rather than route around it. XBRL support matters because financial reports are structured, regulated, and full of the exact numbers analysts want to query. Email support matters because corporate knowledge lives in .msg files that no clean pipeline touches. Video support, with keyframes and transcripts, matters because a growing share of institutional memory is now recorded meetings rather than written memos.

The chart understanding feature deserves separate attention. Charts are where documents hide their conclusions. Converting a bar chart into a table or code and attaching a description is the difference between a retrieval system that finds the slide and one that can answer a question about it.

Why the foundation and the license matter

Docling is MIT-licensed, with individual model licenses handled by their original packages. It runs locally, which the project explicitly frames as support for sensitive data and air-gapped environments. That combination is not incidental. It is the pitch to regulated industries: banks, law firms, defense contractors, and hospitals that cannot ship documents to a hosted API.

The governance choice is also notable. Hosting under LF AI & Data, with IBM Research Zurich as the originator, positions Docling as neutral infrastructure rather than a vendor product. Compare that with the proprietary document AI services from the major cloud providers, which charge per page and keep the parsed output inside their walls.

The integration list reinforces the strategy: LangChain, LlamaIndex, CrewAI, Haystack, plus an MCP server and a docling-serve API mode. Docling is not trying to be the application. It is trying to be the layer every application calls first.

The VLM bet and the GraniteDocling question

Docling supports visual language models, and the README points to GraniteDocling as the reference implementation. The CLI exposes a --pipeline vlm --vlm-model granite_docling flag, so users can swap the classical layout-and-OCR pipeline for a VLM-based one.

This is the most consequential design choice in the project, and it is worth watching closely. Classical document parsing pipelines are fast, cheap, and predictable. VLMs are slower, more expensive, and better at the messy cases: handwriting, unusual layouts, figures with embedded text. Docling’s decision to support both, rather than pick a side, is pragmatic. It also means the project is hedging on which approach wins.

{/* TODO: verify GraniteDocling model size, release date, and benchmark claims — searched GitHub README and IBM Granite materials, did not find authoritative figures in the supplied source */}

The metadata extraction and complex chemistry understanding features listed as “coming soon” point to where this goes next. Molecular structures in particular are a signal that Docling is targeting scientific and pharmaceutical pipelines, where the documents are dense, the formats are hostile, and the cost of a parsing error is high.

What this means for AI builders

Three practical implications.

First, ingestion is consolidating. A year ago, teams stitched together PyPDF, Tesseract, Unstructured, and custom regex. Docling is trying to be the single dependency, and the format coverage is now broad enough that the argument is credible for most enterprise use cases.

Second, the local-execution story is a real differentiator. If you are building for healthcare or finance, the ability to run the full pipeline on-premises without a hosted API call is not a nice-to-have. It is often the difference between a deal and no deal.

Third, watch the VLM pipeline. If GraniteDocling and comparable models get fast and cheap enough, the classical OCR stack becomes a fallback rather than the default. Docling is positioned for that transition either way, which is the smart place to stand.

The project dropped Python 3.9 support in version 2.70.0 and now requires 3.10 or higher, running on macOS, Linux, and Windows across x86_64 and arm64. That is a small detail, but it tells you the maintainers are willing to move forward rather than carry legacy compatibility forever. The technical report is on arXiv at 2408.09869, authored by the Deep Search Team, version 1.0.0, dated August 2024.

The open question is whether Docling can keep pace with the format sprawl it is inviting. Every new parser is another surface to maintain, and the project is now claiming video, patents, scientific articles, financial reports, and email. That is a lot of ground for a foundation-hosted project to hold. The next release will show whether the breadth is sustainable or whether some parsers quietly rot.