LlamaIndex publishes a draft Document Parsing Protocol to standardise how AI reads documents

LlamaIndex, the company behind the LlamaParse document service, has put forward a draft specification describing how parsers should convert documents into text that AI systems can use. Called the Document Parsing Protocol, or DPP, it was introduced in a post dated October 28 on the company's blog and comes with its own website and an open GitHub repository.

What the draft covers

The problem it targets is familiar to anyone building retrieval or agent pipelines: two parsers given the same PDF can produce quite different output for a table spanning two pages, a checkbox on a form or a chart. The DPP site argues that parsers make thousands of such choices and rarely document them, and proposes a shared set of conventions so that parsers, annotators and evaluators interpret a page in the same way.

The guide is labelled working draft 0.1 and advertises 140 rules, although another section of the same site refers to 145 notes. They are grouped by element type: document-level issues such as tables continuing across pages, layout and reading order, headings, inline text and marks such as stamps, signatures or redactions, lists, tables, forms, charts, and images including blueprints, maps and schematics. A final section even addresses chemical structures and sheet music. Three principles run through the text: keep what is printed, encode only the structure the page actually supports, and flag uncertainty instead of guessing. Output is defined as Markdown, with optional extra fields alongside it.

Openness and limits

Code in the project is released under the MIT licence, while the protocol and its examples use Creative Commons CC BY 4.0. The site invites outside contributors to propose rules, submit examples and open pull requests on GitHub, though it does not describe a formal governance or review process. LlamaIndex says the conventions grew out of its ParseBench evaluation work, and the project carries LlamaIndex branding.

Several caveats apply. Some rules are marked as settled only within this draft, others remain open, and the bundled checker is presented as a review aid rather than an automated conformance test. Whether other parsing vendors or model makers adopt DPP will decide if it becomes a shared standard or remains one company's documented house style. For now, it is an early proposal open for comment.