Introducing Agentic Search | Mistral

This post introduces Agentic Search, an enhancement to the Mistral Search Toolkit. It details a multi-step retrieval loop that allows AI models to interact with data using five specific tools: `search`, `open`, `navigate`, `read`, and `grep`. The post provides concrete benchmarks on FinanceBench and OfficeQA Pro, showing improvements in correctness (e.g., 26.7% to 86% on FinanceBench) and reductions in latency (p90 drop from 255s to 154s with navigation) and token usage (e.g., -23.9% for MM 3.5 with navigation). It contrasts this with traditional one-shot RAG, highlighting its limitations with long, dense, or multi-document queries. The core technical contribution is the introduction of these agentic tools for more sophisticated document navigation and information retrieval.