LlamaIndex introduced Extract v2.5 on 2026-10-01 (Adrian Lyjak and Eli Stewart): a new generation of schema-based document extraction agents, with a new agent harness purpose-built for document extraction (LlamaIndex blog).
What shipped
Accuracy improvements are claimed across all three tiers—Cost Effective, Agentic, and Agentic Plus. Advanced Citations (bounding-box grounding of supporting evidence) are improved and now available on Agentic and Agentic Plus; an earlier version had been Agentic Plus only. LlamaIndex says the lifts come with no increase in per-page pricing—higher performance per dollar as vendor framing, with no dollar rates published on the post (LlamaIndex blog).
ExtractBench (vendor-attributed)
On ExtractBench, LlamaIndex reports overall value F1 moving 87.1 → 93.9 (Cost Effective), 89.8 → 95.8 (Agentic), and 95.1 → 96.4 (Agentic Plus), plus grounding scores 46.8 → 80.6 (Agentic) and 46.4 → 82.2 (Agentic Plus). Treat every figure as a LlamaIndex / ExtractBench vendor result, not an independent newsroom measurement or a competitor head-to-head (LlamaIndex blog).
Worked examples on the same bench cover long lists, records that span pages, and scanned forms—still vendor challenge color.
Native spreadsheet mode
Extract v2.5 also adds native spreadsheet extraction: in spreadsheet mode, agents work directly with workbook cells rather than a flattened representation. Enable it in the extraction configuration (LlamaIndex blog).
Who should care
Teams running schema-driven document extraction—especially where citations and confidence scores feed human-in-the-loop review—should start at the Extract v2.5 post. Keep ExtractBench numbers attributed to LlamaIndex, stick to the same-per-page pricing claim without inventing rates, and leave vector-index products for their own desks.
The Campfire
No commentsNobody has pulled up a log by this one yet. Be the first to say what you make of it.
Held for the desk. It appears after a look.