CorpusMap entity map cuts AI agent token use by 57–78%

Researchers at KAIST and Microsoft introduced CorpusMap, a layer between document storage and search agents that builds a map of recurring entities with verified backlinks to source files. On EnterpriseRAG-Bench, token use fell from 206,500 to 88,100 (−57%) and accuracy rose from 62.1% to 73.8%; on WixQA, tokens dropped from 337,200 to 74,500 (−78%).

• EnterpriseRAG-Bench: 206,500 → 88,100 tokens per query, accuracy 62.1% → 73.8%
• WixQA: 337,200 → 74,500 tokens (−78%), accuracy 67.5% → 70.7%
• Indexing 2,819 documents: $4,732 with GPT-5.5 vs $74.65 with GPT-5.6 Luna
• Cheap Luna-built map scores 73.59–74.68 with frontier answering models