LangChain's OpenWiki and evidence graph raise agent accuracy on cross-document questions

LangChain says its open-source wiki compiler, OpenWiki, raised agent accuracy on connected questions from 0.10 with keyword search to 0.52 in a test on a 111-document synthetic insurance library, and cut cost per correct answer from about $0.55 to $0.19.

Citation overlap across repeated runs fell to 0.25 on those questions, which LangChain says an evidence graph layered on the wiki was meant to address.

In a later X post, LangChain says adding its lightweight evidence graph to OpenWiki gives 5x higher accuracy than grep alone, 27% more accuracy than the wiki without the graph, and 2x citation consistency on complex, cross-source questions. It also reports 58% lower cost per correct answer with the graph. These figures come from LangChain's own testing and have not been independently verified.