LlamaIndex 搞了个 ExtractBench，测了 14 个系统，发现长文档提取召回率很低。它家 Agentic Plus 能到 96.1%，比别的都稳。
ExtractBench 基准包含 370 份企业文档和 14 个提取系统。最难的测试是长列表完整性，例如 26,725 行的未认领财产清单。前沿视觉语言模型在长文档上 F1 仅为 8.9%–35.8%，精度高但召回率大幅下降。LlamaIndex 的新方案 Agentic Plus 将长文档迭代处理，在长列表任务上达到 96.1% F1。
The most dangerous document extraction failure isn't a wrong value. It's a missing row that looks li...
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… Y Paper: arxiv.org/pdf/2607.29677 u Your browser does not support the video tag. 🔗 View on Twitter 💬 3 🔄 2 ❤️ 9 👀 1239 📊 4 ⚡
- Jerry Liu08-11 15:14原文