LlamaIndex Drafts Document Parsing Protocol

LlamaIndex on October 28 published a working draft of the Document Parsing Protocol (DPP), a set of 140 rules that spell out what a correctly parsed document page should look like when AI systems convert PDFs, forms and charts into text. The company released version 0.1 as an open draft and asked developers to challenge the rules on GitHub.

Key takeaways

• DPP 0.1 contains 140 rules covering text, tables, forms, charts, diagrams and more, each paired with example pages and expected output.
• The rules grew out of the annotation guidelines LlamaIndex used to label pages for ParseBench, its document parsing benchmark.
• Code is MIT-licensed and the protocol text and examples are under CC BY 4.0; feedback goes through GitHub issues and pull requests.

What the Document Parsing Protocol defines

In a LlamaIndex blog post by Gauthami Polasani and Pierre-Loic Doulcet, the company calls DPP "an answer key for document parsing, with the reasoning behind each hard call." Each rule comes with real example pages shown beside the output a parser is expected to produce.

The project site, parseprotocol.com, organizes the rules into 13 chapters, from conventions and output model to tables, forms, charts, and images and diagrams, plus appendices on equivalent forms and producing ground truth.

How DPP handles tables, forms and charts

Several rules settle long-running disagreements among parsing tools. Tables with multi-row headers are written as HTML, because Markdown cannot show headers that span columns. Empty form fields become "[blank]," and checkboxes are written as "[x]" or "[ ]."

Charts are converted into tables of their underlying numbers, and diagrams become diagram code such as Mermaid where possible. Text that cannot be read is marked "[illegible]," while text hidden inside a PDF is left out. Some variation is allowed: wrapping table rows in thead and tbody tags is acceptable as long as a document uses one form consistently.

Why a shared parsing standard matters for AI agents

The rules originate in the guidelines annotators followed when labeling ParseBench. According to the ParseBench paper, submitted to arXiv on April 9, the benchmark holds about 2,000 human-verified pages of enterprise documents from insurance, finance and government.

That study tested 14 methods, including vision-language models, specialized parsers and LlamaIndex's own LlamaParse, across five dimensions: tables, charts, content faithfulness, semantic formatting and visual grounding. LlamaParse Agentic scored highest at 84.9%, but the authors wrote that "no method is consistently strong across all five dimensions." Without agreed rules for what correct output looks like, such scores are hard to compare across vendors.

Using DPP with Claude and ChatGPT

LlamaIndex offers the protocol as a downloadable Claude skill, which users add under Customize and Skills with code execution turned on, and as a reference file that can be attached to a ChatGPT chat. A walkthrough on the site explains eight parsing decisions on a sample Apple financial statement page.

The GitHub repository keeps one Markdown file per rule. Its code is released under the MIT license, while the protocol and examples use CC BY 4.0 with credit to "Document Parsing Protocol, LlamaIndex."

What's next

LlamaIndex describes 0.1 as a working draft and asks readers to open an issue or pull request with a page and its expected output when a rule fails or conflicts with another. The authors said they want people to "argue with them in the open." The company has not announced a timeline for a later version or named outside partners.

FAQ

What is the Document Parsing Protocol?
It is a proposed standard from LlamaIndex that defines how a document page should be converted into structured text, with 140 rules and example pages in its 0.1 draft.

Is DPP free to use?
Yes. The code is MIT-licensed, and the protocol and examples are published under CC BY 4.0, which requires attribution.

How can developers propose changes?
By opening an issue or pull request on the run-llama/document_parsing_protocol GitHub repository with the page in question and the output they expect.

References

• Introducing the Document Parsing Protocol — LlamaIndex (Oct. 28, 2026)
• run-llama/document_parsing_protocol — GitHub
• ParseBench: A Document Parsing Benchmark for AI Agents — arXiv (Apr. 9, 2026)
• Document Parsing Protocol — parseprotocol.com