Extract V2.6新增最大努力模式
LlamaIndex推出Extract V2.6,为agentic_plus层级新增max effort设置,通过增加计算量并在需要时以更高分辨率检查视觉内容,改进复杂文档的数据提取。公司称,ExtractBench Pro的Value F1从标准Agentic Plus的77.7升至92.9,但不同类别的效果不一,图表和计数任务反而下降。计费随计算需求变化,当额外计算能明显提升质量时,费用可能高于标准agentic_plus任务。
该更新为处理高难度文档且提取错误代价高昂的企业提供了更高质量选项,但LlamaIndex的分项结果表明,迁移前仍应以代表性样本评估质量和可变成本。
Today we’re introducing max effort, a new setting for Extract’s agentic_plus tier. It is designed to deliver our highest-quality extraction and reduce the risk of missing or incorrect data on your most critical and complex documents. Max effort uses variable billing based on the computation required.
As part of our onboarding process for many enterprise customers, we conduct a proof of concept where we evaluate various Extract configurations on samples of the datasets they plan to run in production. An important part of this process is figuring out the optimal tier and schema that will meet the needs of their business while staying within budget. In many cases, the cost that matters more than the price of the extraction service is the cost of a missing or incorrect value. Customers ask, “Can we pay more to unlock even higher quality?”
With max effort on agentic_plus , the answer is now yes. We have developed methods that significantly improve extraction on some of the hardest documents. On ExtractBench Pro, Value F1 rose from 77.7 with standard Agentic Plus to 92.9 with max effort. When max effort is set, billing varies according to the complexity of the extraction and the computation required. That additional quality comes with more computation and can cost more than a standard agentic_plus job.
#LlamaIndex#LlamaIndex Extract V2.6#ExtractBench Pro#LlamaCloud