Docling prepares complex documents locally for AI

What this is about

Docling is an open-source document-processing tool. It reads formats including PDF, DOCX, PPTX, XLSX, HTML, EPUB, images, audio, video, and email, then converts their content into a shared structure. This is useful when documents need to be prepared for search, retrieval-augmented generation (RAG), data analysis, or AI agents.

The project was started by IBM Research and is now hosted by LF AI & Data. It released version 2.120.1 on August 14, 2026. Its practical value is not another chat window but the harder preparation step: Docling attempts to preserve reading order, tables, formulas, code blocks, images, and page layout instead of extracting only an unstructured stream of text.

What Docling actually does

Docling can run locally through a command-line interface or as a Python library. A basic command such as docling file.pdf produces structured output such as Markdown. Applications can also use DoclingDocument as a common representation and export content to Markdown, HTML, WebVTT, DocTags, or lossless JSON.

The project provides OCR for scanned documents. For demanding pages, users can select a pipeline based on vision-language models. Its documentation also lists integrations with LangChain, LlamaIndex, CrewAI, and Haystack, an MCP server for agents, and docling-serve as an API service. The same conversion logic can therefore support local experiments, enterprise services, or automated knowledge pipelines.

The current feature list goes beyond conventional office documents. Docling also handles audio with speech recognition, video with transcripts and representative keyframes, email, OpenDocument files, and XBRL financial reports. For charts, the project describes functions that can turn bar, pie, and line charts into tables or code and add descriptions.

Why it matters

Many AI applications fail first because of poor input data, not because of the language model. If columns are mixed, tables are destroyed, or footnotes are assigned to the wrong passage, downstream search is built on a faulty foundation. Docling makes this preparation a visible and programmable step that teams can inspect.

Local execution is especially relevant for teams working with confidential contracts, internal reports, or technical records. According to the project documentation, Docling can run in air-gapped environments. That reduces unnecessary transfers to external services, but it does not replace a privacy review: any OCR or vision models, logs, temporary files, and connected systems must also be controlled.

The project uses the MIT license, while individual integrated models may have different terms. With more than 64,000 GitHub stars on August 16, 2026, Docling has substantial visibility. Stars do not prove accuracy or suitability for a particular document collection. The decisive step is testing it with real files from the intended workflow.

In plain language

Docling works like a careful mailroom. It does not merely open an envelope and dump every word into a box. It recognizes the address, attachments, tables, and page order, places them in labeled trays, and gives the next department an organized case file. If the original is smudged or unusually arranged, the mailroom can still sort it incorrectly, so sampling is necessary.

A practical example

A midsize machinery company has 2,000 maintenance reports containing tables, scanned drawings, and handwritten additions. The team wants to build an internal search system that sends service technicians to the correct page.

It first selects 100 typical and 20 particularly difficult reports as a test set. Docling converts them locally into structured documents. A human then checks reading order, table cells, page references, and OCR results. Only after defined minimum quality thresholds are met are sections added to the search index with document IDs and page numbers. The later AI answer links back to the original. Classes of documents that fail are not silently accepted; they are handled with another pipeline or reviewed manually.

Scope and limits

Docling is a capable preparation layer, not a truth checker.

• Complex scans, handwriting, nested tables, and unusual layouts can be parsed incorrectly. A representative gold set and ongoing samples remain necessary.
• Local operation does not automatically mean low effort. Models need memory and compute, while an API service, upgrades, and queues need an operating plan.
• Structured output does not prevent hallucinations in a downstream language model. Citations, access controls, and answer testing remain separate tasks.
• The core MIT license does not automatically cover every model used with it. Teams must review model licenses and data rights separately.

The most sensible first test is small: convert ten simple and ten difficult documents, compare the outputs manually, and define measurable stop criteria. Only then is an integration into RAG or an agent worth the effort.

SEO & GEO keywords

Docling, IBM Research, LF AI & Data, document processing, PDF parsing, OCR, RAG, local AI, MCP server, DoclingDocument, GenAI data pipeline, open-source AI