Report
ExtractBenchPro launches to test complex document extraction
The team behind LlamaParse says it tests document-specific rules, jargon and values inferred from multiple sources.
TLDR
The team behind LlamaParse says ExtractBenchPro draws on difficult customer documents, including trial balances and well logs. Its headline metric is the share of documents with a perfect value F1 score. The team reports that Codex with Astra reached about 58% Value F1 at almost $1 per page; it puts its own agentic modes at 20–40%, at prices it describes as feasible for production use.
Combined views
730
3 Sources, first seen ago
