docling

Docling turns messy documents into clean, structured data for AI. Feed it PDFs, Word docs, PowerPoints, spreadsheets, images, HTML, email, even audio and video, and it produces Markdown, HTML, or lossless JSON with the layout, reading order, tables, formulas, and code preserved. Started at IBM Research Zurich and now hosted by the LF AI and Data Foundation, it is MIT licensed and free.

It is a Python library (3.10 or later) that runs locally, which is the point: documents never leave your machine, so it works in air-gapped or sensitive environments. It plugs into RAG pipelines with integrations for LangChain, LlamaIndex, Crew AI, and Haystack, and there is a CLI, an API server, and an MCP server for agents. Heavy parsing with vision models wants a GPU.

Solo to large teams: free, and strong at the RAG-prep job. It is less a raw OCR engine and more a document-to-structured-data framework, the same job as Textract's Analyze Document. Teams that do not want to run it can now buy it: IBM sells Docling for watsonx, a managed version of the same open stack, from $4 per 1,000 pages.

The catch is that quality depends on the document. Layout and table understanding is good, but complex or badly scanned PDFs still trip it up. It is a framework, so you own the pipeline and the tuning. Powerful, but not a fire-and-forget API.