ExtractBench: Document Extraction Benchmark
ExtractBench is an open benchmark by LlamaIndex for schema-guided extraction from enterprise documents, including scans and handwriting. Systems fill the same JSON Schema from the same document and are evaluated on value accuracy, record completeness, source grounding, and cost. The benchmark provides a public dataset, reproducible evaluation code, a research paper, and an interactive leaderboard.
Dataset and evaluation
- Download the ExtractBench dataset on Hugging Face
- Run the open-source ExtractBench evaluation
- Research paper