What changed
ExtractBench evaluates 370 enterprise documents spanning 4,869 pages, eight business domains, and 67 document types. It measures value accuracy, record completeness, page- and word-level grounding, and cost, exposing truncation on long repeated-record documents.
Why it matters
A system can extract several fields correctly while silently dropping most rows from a schedule, table, claim list, or transaction record.
What to do
Add record-completeness and evidence-location checks to document extraction tests, and evaluate short forms separately from long repeated-record documents before choosing a model or agent.
Caveat
LlamaIndex researchers created the benchmark and their commercial extractor ranked first. Some ground truth uses system agreement, and the advertised public artifacts were not reliably accessible when verified.