LlamaIndex wants parser vendors, benchmark builders and annotators to agree on what a correctly parsed page looks like. On October 28 the company released the Document Parsing Protocol, or DPP, as a public working draft numbered 0.1. It sets out 140 rules for converting document pages into structured Markdown, spanning running text, tables, forms, charts and diagrams, and pairs each rule with sample pages and the output the rule expects.
The problem it targets is easy to miss. Two parsing tools can recover every word and number on a page and still disagree on structure: whether aligned text is really a table, whether a two-tier header gets merged into one line, whether a table that breaks across pages counts once or twice, and what to do with a flowchart. LlamaIndex argues that these disagreements travel downstream, because an AI agent reasons only over what the parser hands it, and no common reference currently exists for judging which structural choice is right.
The company's own illustration is a sales table with years across the top and quarters beneath them. A parser that fuses the two tiers into labels such as "2024 Q2" may still let an agent answer correctly, yet DPP counts that output as non-conforming, since the label never appeared on the page and each tool would coin it differently. The protocol instead keeps every printed header tier on its own row with exact column spans. Plain Markdown pipe tables cannot express spans, so DPP calls for HTML table markup wherever Markdown lacks the syntax.
Other rules aim to keep information that parsers commonly drop. An unanswered form field is written as [blank], so it cannot be mistaken for a field the tool missed, and checkboxes survive as [x] or [ ]. Charts become data tables, each category getting a row and each series a column, and values estimated from bars or points are limited to the precision the axis supports. Diagrams go into a language such as Mermaid when labels and arrow directions can be kept, text hidden in the file but invisible on the page is left out, and visible text that cannot be read is tagged [illegible].
DPP grew out of the annotation conventions LlamaIndex used to label pages for ParseBench, its document-parsing benchmark. The company presents it as a shared answer key for anyone who builds benchmarks, labels training data or reviews parser output, so that two evaluations stop accepting different answers for the same page. The draft also spells out tolerated variation: wrapping table rows in thead and tbody elements or leaving them out both conform, provided a document sticks to one form.
The project is published in the open. Code in the run-llama/document_parsing_protocol repository on GitHub is MIT-licensed, while the protocol text and its examples carry a CC BY 4.0 license that permits commercial reuse with attribution. A companion site, parseprotocol.com, places each rule beside its expected output, offers a Claude skill and a ChatGPT reference file for testing the rules on a page image, and walks through Apple's FY2024 statement of operations. LlamaIndex is asking for issues and pull requests wherever a rule fails on real documents or two rules contradict each other.
The release lands days after LlamaIndex introduced its OCRia-1 document model on October 22, and it sits close to the company's commercial parsing product, LlamaParse. That makes neutrality the open question. The site itself says a "settled" rule is settled only within this draft rather than ratified by any outside body, and that the bundled checker is review guidance, not an automated conformance suite. Whether rival parser vendors and benchmark maintainers adopt DPP, or answer with competing conventions, will decide if it becomes a shared standard or stays one company's labeling guide.