Primary sourceLlamaIndex
View sourceEvent ·Single source
ExtractBench:14个文档提取系统各有感知盲区
ExtractBench测试了14个文档提取系统,使用1950年代监管文件、手填税表和手机拍摄等非数字化文档。Codex在扫描和手写场景准确率超93%,但处理旋转或纯图像页面时降至约80%。专用API表现相反,旋转和手写识别良好,扫描准确率为81%。Gemini 3.5 Flash从干净PDF的88.6%跌至扫描件的71.1%。LlamaIndex的Agentic Plus是唯一无盲区系统,在旋转、扫描和手写三项上分别达到95.9%、93.9%和93.8%。
Current read
LlamaIndex测了14个文档提取系统,Codex、Gemini在扫描件上都掉链子,只有Agentic Plus三项全在93%以上。
2 sources63° AI heatLast updated 2026/8/13 14:05:11
Evidence
2 sourcesRelated source 1Jerry Liu
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.