Skip to content
AI.info

The Pulse

LlamaParse gets two OCRia-1.5 models for document parsing

LlamaIndex has put two OCRia-1.5 document parsing models into its LlamaParse service. The company reports higher scores for tables, charts, forms and visual grounding on its own benchmark.

LlamaParse gets two OCRia-1.5 models for document parsing

LlamaParse now runs two OCRia-1.5 document-parsing models, with a smaller model powering its Cost Effective tier and a larger one running Agentic. The models are available in the service, LlamaIndex said in a product post, and the company reports improved scores on its own benchmark for tables, charts, forms and visual grounding.

Two models and credit rates

The smaller OCRia-1.5-2B has half the parameters of the earlier OCRia-1 model. The larger OCRia-1.5-4B is the upgraded model for the Agentic tier.

Cost Effective runs at 3 credits per page, while Agentic runs at 10 credits per page. LlamaIndex positions the lower-cost tier for high-volume parsing and says Agentic remains the more accurate choice for complex tables and charts.

The company says the new models are more accurate than OCRia-1 on complex tables, charts and forms, and place bounding boxes more precisely. It also reports that Cost Effective improved across all six dimensions of ParseBench-2, its document-parsing benchmark.

Tables, charts and forms

Complex tables were a focus of the update. In a financial statement, for example, a heading can cover several columns; dropping it can leave the figures intact but remove the period to which they apply.

In its tests, LlamaIndex says table scores rose from 86.5 to 90.1 for Agentic and from 83.2 to 87.5 for Cost Effective. Gains were larger on tables with headers spanning multiple columns.

The company reports that chart scores rose from 74.9 to 76.8 for Agentic and from 67.3 to 73.0 for Cost Effective. Forms scores rose by 2.6 points and 4.2 points respectively; that measure checks whether values stay with their fields and checkboxes retain their state.

Locating content on a page

LlamaParse bounding boxes can be used to highlight the source behind an answer. A box that points to neighboring text or leaves part of a table outside the highlight can make it harder to check that answer against the page.

With the new models, each form field gets its own box, and separate tables on a page are boxed individually, LlamaIndex says. The company says visual grounding scores rose from 77.0 to 88.5 for Cost Effective and by 4.1 points for Agentic.

How the benchmark works

ParseBench-2 evaluates text, tables, formatting, charts, layout and forms. Its document corpus expanded from 2.1k to 4.3k cases from the previous benchmark, and its scoring uses severity levels assigned to types of parsing error.

The chart evaluation compares output with reference tables covering 36,958 values across 1,025 charts. For forms, the core dataset contains 500 pages from 244 government form templates across 30 countries, including 450 with synthetic typed values and 50 blank pages.

Sources

  1. LlamaIndex
  2. LlamaIndex

Companies in this story

Related stories

Explore

More articles