LlamaIndex has released ParseBench-2, an open-source benchmark that quantifies the accuracy of artificial intelligence when extracting data from complex documents, and a accompanying family of parsing models called OCRia-1.5. The new benchmark includes more than 4,300 document pages and focuses on the severity of production errors, assigning weights to more than 330 error types categorized from critical to harmless. LlamaIndex states this approach aligns metric scores more closely with human judgment than previous deterministic or LLM-based benchmarks.

The released OCRia-1.5 models now power all modes in the LlamaParse tool, including those for parsing bounding boxes in messy layouts, tables, and charts. Initial results show LlamaParse achieved an overall score of 90.1 on the new benchmark, while the Claude Opus 5.5 model scored 92.7 in text parsing compared to 90.3 for LlamaParse. The ParseBench series is used by companies including Nvidia, Cohere, and Perplexity and currently records more than 50,000 monthly downloads.