Agentic Search
Agentic Search is an information retrieval paradigm in which AI agents autonomously plan and execute search processes. Unlike traditional fixed-query retrieval, agents dynamically determine search strategies, keyword selection, retrieval timing, and result filtering based on task objectives, and can perform multi-round iterative searches. This paradigm typically combines reasoning capabilities with external tool invocation to handle complex, multi-step information acquisition needs, and is widely applied in knowledge augmentation and context construction for agent systems.
Timeline (last 90 days)
Agentic Search tools like Tavily first understand the question, decompose it into sub-questions, select the best data source for each sub-question, then extract, perform vector retrieval, score and filter, and finally return structured JSON.
Products such as OpenAI's Operator and Google's Agentic Search are also advancing similar asynchronous Agent capabilities.
Mistral officially launched Agentic Search, enabling large models to browse through documents like a human rather than mechanically performing a single vector retrieval.
Mistral launched Agentic Search, empowering agents with the ability to locate and browse lengthy and complex documents across multiple rounds through five retrieval tools: Search, Open, Navigate, Read, and Grab.
Mistral Agentic Search improved accuracy on the Finance Bench earnings report evaluation from 26.7% to 86%, reduced P90 latency by 39.6%, and decreased token consumption by approximately one third.
All Facts (6)
Agentic Search的三种失败模式包括：Agent不调用任何工具、Agent调用错误的工具
85%UnverifiedAgentic Search tools like Tavily first understand the question, decompose it into sub-questions, select the best data source for each sub-question, then extract, perform vector retrieval, score and filter, and finally return structured JSON.
50%UnverifiedProducts such as OpenAI's Operator and Google's Agentic Search are also advancing similar asynchronous Agent capabilities.
50%UnverifiedMistral Agentic Search improved accuracy on the Finance Bench earnings report evaluation from 26.7% to 86%, reduced P90 latency by 39.6%, and decreased token consumption by approximately one third.
50%UnverifiedMistral officially launched Agentic Search, enabling large models to browse through documents like a human rather than mechanically performing a single vector retrieval.
50%UnverifiedMistral launched Agentic Search, empowering agents with the ability to locate and browse lengthy and complex documents across multiple rounds through five retrieval tools: Search, Open, Navigate, Read, and Grab.
50%