LlamaIndex 🦙(@llama_index)

The most dangerous document extraction failure isn't a wrong value. It's a missing row that looks li...

8.5内容质量

TL;DR · AI 摘要

LlamaIndex发布ExtractBench基准测试，揭示长文档提取中缺失行的严重问题，并推出Agentic Plus系统实现96.1% F1分数。

核心要点

  • ExtractBench包含26,725行未申报财产列表等测试数据，验证长列表完整性挑战
  • Agentic Plus系统通过迭代处理长文档，F1分数达96.1%，优于其他系统
  • 前沿VLMs处理长文档时召回率低，F1分数仅8.9–35.8%

结构提纲

按章节快速跳转。

  1. §引言

    指出文档提取中缺失行比错误值更危险，引出ExtractBench基准测试

  2. ·ExtractBench基准测试

    包含370个企业文档和14个系统，重点测试长列表完整性

  3. ›长列表完整性挑战

    测试数据包含26,725行未申报财产列表等极端案例

  4. ·VLMs局限性

    前沿视觉语言模型在长文档处理中召回率低，F1分数8.9–35.8%

  5. ·Agentic Plus解决方案

    通过迭代处理长文档，实现96.1% F1分数且性能随文档长度增长保持稳定

思维导图

用一张图看清主题之间的关系。

查看大纲文本（无障碍 / 无 JS 友好）
  • 文档提取挑战与解决方案
    • ExtractBench基准测试
      • 测试数据：26,725行未申报财产列表
      • 14个系统对比测试
    • Agentic Plus系统
      • 迭代处理机制
      • 96.1% F1分数

金句 / Highlights

值得收藏与分享的关键句。

#文档提取#LlamaIndex#ExtractBench#Agentic Plus
打开原文

LlamaIndex 🦙 on X: "The most dangerous document extraction failure isn't a wrong value. It's a missing row that looks like nothing is wrong. We released ExtractBench yesterday: 370 enterprise docs, 14 systems. The hardest test: long-list completeness. An unclaimed-property list with 26,725 rows. A https://t.co/NgY4ZHFirf" / X

LlamaIndex 🦙

@llama_index

The most dangerous document extraction failure isn't a wrong value. It's a missing row that looks like nothing is wrong. We released ExtractBench yesterday: 370 enterprise docs, 14 systems. The hardest test: long-list completeness. An unclaimed-property list with 26,725 rows. A creditor matrix with 8,624 records. A 13F with 3,063 holdings. Frontier VLMs don't misread these docs, they abandon them. Precision stays high, recall collapses: 8.9–35.8% F1 on the longest documents. Every row they return looks correct, so spot checks pass while most of the document never came back. Our new Extract tier, Agentic Plus, processes long docs iteratively instead of one pass: 96.1% F1 on long-list tasks, and the only system that holds flat as docs get longer. Learn more about ExtractBench below👇 Blog:

llamaindex.ai/blog/introduci…

Paper:

arxiv.org/pdf/2607.29677

$

00:00

/$

4:20 PM · Aug 12, 2026

3.4K

Views

3

4

24

8