ExtractBench: A Benchmark for Schema-Guided Enterprise Document Extraction

Researchers released ExtractBench, a new benchmark for enterprise document extraction systems.

It tests schema-guided extraction with source evidence across 370 documents, 4,869 pages, 8 business domains, and 67 document types.

The benchmark measures value accuracy, record completeness, grounding, and cost to better reflect real-world workflow constraints.

Commercial vision-language models performed well on shorter documents but struggled with long record lists, while coding agents were more accurate but far more expensive.

LlamaExtract Agentic Plus ranked best across the benchmark’s main metrics.