Mistral gives document search an agentic loop

Mistral says its new Agentic Search retrieval loop lifts FinanceBench correctness from 26.7% to 86%—more than 3×—and raises table-heavy OfficeQA Pro from 6.3% to 51.9%, while cutting p90 latency by up to 39.6% and token use by up to one-third. The system adds search, open, navigate, read, and grep tools on top of an existing index, so the model can inspect long filings and tables instead of trusting one retrieved chunk. Teams building document agents should file this beside the archive’s search-API cost benchmark: retrieval quality is increasingly a tooling and workflow question, not just a model-selection question.