What it is
Docling converts documents into AI-ready text. It reads PDF, DOCX, PPTX, XLSX, HTML, EPUB, email (EML and MSG), LaTeX and images, and lately audio and video through speech recognition, and outputs Markdown, HTML or a lossless JSON document model that keeps headings, reading order and tables.
PDFs and images go through local machine-learning models for layout detection, table structure (TableFormer) and OCR, with an optional vision-language mode using IBM's 258M-parameter GraniteDocling. It includes chunking for RAG, an API server (docling-serve), an MCP server, and integrations with LangChain, LlamaIndex, CrewAI and Haystack.
It started at IBM Research and is hosted by the LF AI & Data Foundation. Releases are frequent: v2.142.1 shipped on 23 October 2026.
Our review
Docling is the open-source parser to beat for PDFs with real tables, and its weak spot is now security rather than quality. It has about 69,400 GitHub stars, an MIT licence and a neutral foundation home, and Kanopy Labs' 2026 comparison found its table extraction stronger than Unstructured's on complex PDFs.
The output is what makes it useful for RAG. Docling keeps reading order, heading levels and table structure in a typed document model, so chunks carry their section context instead of arriving as loose text. It also runs locally: the FAQ says it uses no remote service and can work in an air-gapped network once the model weights are copied in.
Reading so many formats creates a large attack surface. The project has published 28 GitHub security advisories, 11 of them on 23 October 2026 alone, mostly small files that exhaust memory or CPU (oversized archive members in DOCX, XLSX and EPUB, unbounded LaTeX table spans) plus a TikZ check bypass that allowed file read and write. A high-severity advisory on 1 October covered arbitrary file read and write. The latest batch is fixed in 2.142.1, and a service parsing uploads from strangers should pin to it and isolate the worker.
Speed and install size are the practical limits. The GPU benchmark page reports roughly 1 to 8 pages a second per machine across an NVIDIA L40S, RTX 5090 and RTX 5070, depending on the pipeline and whether OCR runs, so a million-page archive is about three days of work at four pages a second. The ML stack arrives as hard dependencies, and reviewers report weaker results on handwriting and phone photos, where the vision-language pipeline or a hosted parser is the usual fallback.
Where it earns its place
A platform engineer at a 50-person legal tech startup
Ingest contracts and court filings full of tables into a RAG index
Run docling-serve on one GPU node, send each upload through it and store the JSON model and chunker output. No licence cost; at a few pages a second, a 200,000-page backlog is under a day of GPU time, and the parser must be isolated and patched because it reads untrusted files.
A compliance manager at a regional bank
Process internal reports without sending documents to a third-party API
Docling runs air-gapped once its model weights are installed, so documents never leave the network and no vendor data-processing review is needed. The trade-off is owning the hardware and keeping up with frequent security releases.
A data scientist building a research corpus
Convert arXiv PDFs, LaTeX sources and EPUBs into Markdown for fine-tuning
pip install docling and run the CLI over a folder, switching to the GraniteDocling vision pipeline for pages heavy with formulas. Free; CPU-only runs are slower, so a large corpus justifies a GPU.
Before you commit
- Setup
- pip install docling on Python 3.10 or later (3.9 support ended in 2.70.0); the first run downloads layout, table and OCR weights. One CLI command converts a PDF, and docling-serve provides an API server.
- Integrations
- Loaders for LangChain, LlamaIndex, CrewAI and Haystack; an MCP server for agents; export to Markdown, HTML, DocTags and lossless JSON.
- Data and privacy
- Runs locally and calls no remote service by default, according to the FAQ; remote image fetching and remote OCR are opt-in. MIT licence with no commercial limits. 28 published security advisories, the latest fixed in 2.142.1.
- Cost at scale
- The software is free and the cost is compute. The project's benchmarks report roughly 1 to 8 pages a second per machine, so heavy ingestion means a dedicated GPU node such as an AWS g6e.2xlarge with an L40S. The maintainers sell no paid edition.
- Limits
- About 900 open issues and pull requests; text styles such as bold are not extracted from PDFs; handwriting and camera photos remain weak; parsing untrusted files needs sandboxing.
What's good
- Strong table and reading-order recovery
- Runs fully offline with local models
- Lossless JSON model plus Markdown and HTML
- MIT licence under a neutral foundation
What's not
- 28 security advisories; patching is mandatory for untrusted files
- Heavy ML dependencies
- A few pages a second per GPU
- Weaker on handwriting and photos
Best for
Teams building self-hosted RAG over PDFs and Office documents with complex tables.
Alternatives
Docling: common questions
- Is Docling worth it?
- Docling is the open-source parser that gets tables right, and the one whose security releases you cannot afford to skip. We rate it 4.5 out of 5.
- How much does Docling cost?
- Free and open source under the MIT license.
- What are the downsides of Docling?
- 28 security advisories; patching is mandatory for untrusted files; heavy ML dependencies; a few pages a second per GPU; weaker on handwriting and photos.
- Who is Docling best for?
- Teams building self-hosted RAG over PDFs and Office documents with complex tables.
- How much does Docling cost for a team or in production?
- The software is free and the cost is compute. The project's benchmarks report roughly 1 to 8 pages a second per machine, so heavy ingestion means a dedicated GPU node such as an AWS g6e.2xlarge with an L40S. The maintainers sell no paid edition.
- Is Docling safe for company data?
- Runs locally and calls no remote service by default, according to the FAQ; remote image fetching and remote OCR are opt-in. MIT licence with no commercial limits. 28 published security advisories, the latest fixed in 2.142.1.
- What are the best Docling alternatives?
- The closest alternatives we have reviewed are Unstructured, LlamaIndex and Firecrawl.
Answers checked Oct 23, 2026.
Make Docling? Show this rating on your site
