A parser can capture every number on a page and still give an AI agent the wrong answer.
LlamaIndex published the Document Parsing Protocol on October 28. The announcement introduces working draft 0.1: 140 rules covering text, tables, forms, charts, and diagrams.
The proposal focuses on how a document’s meaning survives conversion into structured text. Its examples pair source pages with expected output and explanations of difficult choices.
The problem sits between OCR and the answer
Optical character recognition can recover words without preserving their relationships. In a table, a parent heading may determine which year or business unit a number belongs to. Losing that heading can change the answer without changing the number.
DPP proposes conventions for making these choices explicit. LlamaIndex is asking for contributions and counterexamples. The working draft is a vendor-led proposal; its release does not establish industry adoption or compatibility between existing parsers.
Try a deliberately ambiguous table
For an original test example, give a table with two columns labeled “Q2.” Place one under “2025” and the other under “2026.” Use values of 12 and 18, respectively. Ask for the 2026 Q2 value.
The answer should be 18, but answer correctness is only one part of the evaluation. Inspect whether the intermediate representation keeps each Q2 column attached to the correct year. A flattened output with two identical Q2 headings discards that relationship.
Then swap the two years while keeping the numbers in place. A parser-and-agent workflow that follows the hierarchy should change its answer. Record both the structured output and final answer so a lucky response cannot conceal a structural failure.
These are synthetic test instructions, not results from a parser comparison. They let a team evaluate document fidelity before relying on an agent’s final response.
Check attribution before reusing the protocol
The public repository separates its licensing: software code uses MIT, while protocol text and examples use CC BY 4.0. Its notices also identify material with separate status, including third-party financial statements.
If you adapt the rules or examples, preserve the applicable attribution and exceptions. A single repository license should not be assumed to cover every asset in the same way.
For the model side of document extraction, see our WeVisDoc coverage. DPP addresses representation conventions, which are a separate question from any one model’s accuracy.
