Deep Dive into Long-List Completeness: Evaluating Extraction Completeness on Long Documents

ExtractBench is an open benchmark for schema-guided document extraction. This deep dive looks at long-list completeness: how extraction systems handle multi-page, record-heavy enterprise files. Every commercial VLM we tested scores well on short documents, then falls below 35% recall on documents past 50 pages, while precision stays high. The values that come back are correct. Most of the list doesn't come back at all. ExtractBench scores long-list completeness with deterministic Hungarian alignment for record arrays and no LLM judge, so recall is measured exactly across complex schemas and