LlamaIndex launches Document Parsing Protocol: 140 rules to standardize OCR output for agents
llama_index · x · 2026-10-29
LlamaIndex founder Jerry Liu introduced the Document Parsing Protocol (DPP), a working draft that tackles a chronic problem in document parsing: hundreds of OCR models and tools parse the same document in different ways.
The problem: Some parsing errors are obvious (hallucinated text, misaligned tables), but many judgment calls are ambiguous — is borderless whitespace really a table? Should a two-header-row table preserve both? Parsers decide differently, shaping what your agent understands, and existing benchmarks inherently reward one style.
What DPP does: A curated set of 140 rules covering text, tables, forms, charts, and diagrams — even chemical structures and music notation. Each rule comes with example pages and expected outputs. For instance, fill-in fields and checkboxes are key-value content that must be rendered as key-value lists, not tables. The announcement page walks through a US 1040 tax form showing grouped boxes, empty boxes, printed numbers, and margin notes.
The draft is live for feedback, with a Claude skill and ChatGPT reference file so you can test it on your messiest pages.
Related event: LlamaIndex Unveils Document Parsing Protocol Draft(2 posts)→
More from coding & agent
- Retweeted thesis: Self-updating docs are an 'infinite market' like AWS — handotdev · 2026-10-29
- Vicki Boykis on how to learn ML in 2026: self-taught path and working with agents — vboykis · 2026-10-29
- Coinbase launches global AI trading competition with $2M prize pool — MurrLincoln · 2026-10-29
- vboykis: an agent eagerly helped me set an env var wrong and broke my results — vboykis · 2026-10-29
- Video pro: anyone claiming AI edits videos end to end is lying — juliey4 · 2026-10-29
- One Prompt Batch-Audits and Edits 188 Videos, Author Calls Anti-AI Like Anti-Electricity — yungcontent · 2026-10-29