Thesis: the release’s central change is combining higher reported extraction accuracy with stronger source-level grounding, rather than improving text capture alone. Mechanism: LlamaIndex says it introduced a document-extraction agent harness using structural reasoning, intermediate representations, purpose-built tools, and adaptive inspection; for Agentic and Agentic Plus, Advanced Citations return bounding boxes for evidence and confidence scores can help flag values for review.
Observed facts: LlamaIndex reports ExtractBench value F1 rising from 87.1 to 93.9 for Cost Effective, 89.8 to 95.8 for Agentic, and 95.1 to 96.4 for Agentic Plus, with no per-page price increase. It also says grounding scores rose from 46.8 to 80.6 for Agentic and 46.4 to 82.2 for Agentic Plus. In a 17-page filing example, the company says Cost Effective v2.5 returned all 238 holdings, compared with 87 under v2.0. Why now: document extraction is often used to feed downstream databases and review workflows, where an omitted record or an unsupported value can matter as much as a transcription error; the update explicitly targets those failure modes.
First implication: if the reported gains hold on customer workloads, teams may be able to automate more schema-bound extraction while preserving a human verification path for uncertain fields. Evidence boundary: all scores and examples here come from LlamaIndex’s product announcement and its ExtractBench framing; no independent replication, dataset details sufficient to generalize, or customer-side validation is supplied in the source. First watch condition: independent tests on varied real-world documents should compare both field accuracy and citation quality at equivalent per-page costs before users treat the published gains as general performance guarantees.