Introducing ExtractBench: The Most Comprehensive Benchmark for Data Extraction from Enterprise Docs

Industry Trends and Evaluation Challenges - Enterprise workflows increasingly rely on agents to process unstructured documents, demanding complete, correct, and traceable extraction outputs for meaningful actions[1] - Traditional document extraction benchmarks fail to keep pace with rapid model evolution, typically providing only a single aggregate score over a narrow document set without revealing specific failure causes[1][2][3] Benchmark Performance and System Comparison - ExtractBench evaluates schema extraction in production settings, encompassing 370 enterprise documents across 67 document categories and 8 business domains[2] - Evaluations across 14 systems—including visual language models, open-source pipelines, specialized APIs, and coding agents—reveal that vision model performance collapses on very long documents, and neither vision models nor coding agents provide out-of-the-box visual grounding[4] Company Performance and Technical Advantages - The Llama Extract Agentic+ system achieved the highest overall score of 95.6%, and is the only system scoring above 90 points on long documents and difficult document scans[5] - The Llama Extract Agentic+ system delivers its performance at a cost that is one-third of the next best system[5]