ParseBench brings 2,000 enterprise pages and 167K OCR rules to Kaggle

What the tweet revealed

LlamaIndex put a concrete shape around enterprise OCR evaluation with this line: The first document OCR benchmark built for AI agents — 2,000 enterprise pages, 167K+ test rules, 5 dimensions that actually break downstream agents. The tweet also says the benchmark compares 14 methods, including GPT-5 Mini, Gemini 3, Textract, and LlamaParse, and that the leaderboard is live on Kaggle.

The LlamaIndex account typically posts document-parsing and agent infrastructure updates, so this is squarely in its core beat. The signal is strong because the tweet is not marketing OCR in the abstract. It defines dataset size, evaluation breadth, and downstream failure modes in a way most benchmark launch posts do not.

What the linked post adds

The companion blog post explains why this benchmark exists. Enterprise files are messy: insurance filings, financial reports, contracts, and regulatory submissions carry tables, footnotes, charts, formatting quirks, and visual grounding problems that basic OCR scores often miss. LlamaIndex argues that the standard for AI agents is no longer “human-readable enough,” but reliable enough for an agent to take action without silently misreading a cell, value, or header.

The post says ParseBench evaluates roughly 2,000 human-verified enterprise document pages with more than 167,000 test rules across five dimensions: tables, charts, content faithfulness, semantic formatting, and visual grounding. It also gives one of the more useful contextual comparisons in the piece: even OmniDocBench, described as the most diverse OCR benchmark available, draws only 6% of its pages from enterprise content. LlamaIndex, Kaggle, and the open community are also publishing the dataset, code, and paper, which gives builders something better than a black-box leaderboard screenshot.

What to watch next

The next step is whether teams actually use ParseBench to choose parsers for agents in finance, insurance, and legal workflows, and whether the promised end-to-end agent evaluation arrives. If the benchmark becomes a reference point for procurement and model routing, it could matter more than a generic OCR leaderboard because it tests the kinds of failures that damage real business automation.

Sources: X source tweet · LlamaIndex ParseBench blog · ParseBench Kaggle leaderboard · ParseBench paper