OpenWiki and Graphs Enhance Agent Accuracy in Knowledge Retrieval

LangChain introduces OpenWiki and evidence graphs to improve agent memory and retrieval accuracy. Testing on a synthetic dataset of 111 insurance documents shows OpenWiki raised search accuracy from 0.10 to 0.52, and the addition of evidence graphs further lifted it to 0.58. By avoiding the reconstruction of cross-document relationships on each query, wiki reduces per-query cost from 0.55 to 0.19 dollars. Compiling 111 documents into a wiki costs approximately 4.15 dollars, representing about one-tenth of the original size; adding 17 subsequent documents costs about 0.09 dollars. The technology addresses navigation of distributed knowledge in policies and manuals, though challenges remain in ensuring reproducible evidence paths.

Written with AI from the reporting · updated 10/28/2026