Agentic RAG
Agentic RAG is an AI architectural paradigm that combines agent mechanisms with Retrieval-Augmented Generation (RAG) technology. It encapsulates retrieval operations as callable tools, granting large language models autonomous planning, multi-step reasoning, and dynamic iteration capabilities. This allows the model to independently determine when and what to retrieve based on task requirements, then verify and integrate results, enabling it to handle multi-hop question answering and long-chain reasoning tasks more complex than those addressed by traditional RAG.
Core Facts
The technical evolution of RAG has gone through three stages: Naive RAG, Advanced RAG, and Modular RAG.
Agentic RAG grants the model autonomous planning capability, enabling it to assess whether information is sufficient, decide whether to perform multi-round retrieval, choose which tools or data sources to invoke, and carry out self-reflection and iterative correction.
Traditional RAG follows a unidirectional, fixed, one-shot process, whereas Agentic RAG is a cyclically iterative, multi-step execution process.
Timeline (last 90 days)
In multi-hop questions, answers are scattered across multiple documents and require finding clue A before knowing to look for clue B; the agent loop transforms retrieval from a one-time action into a repeatedly executable tool call.
In the Google FRAMES benchmark, a simple agent loop (capable of reading retrieved results and searching again) achieved an accuracy of 92.7%.
The three-agent collaborative team doubled system latency without delivering any improvement in accuracy.
Architectural complexity should be matched to task benefit; when the candidate space is closed and finite, multi-agent negotiation and verification mechanisms may only increase costs without touching the accuracy bottleneck.
On classification tasks with a closed vocabulary, stacking agent roles does not automatically yield accuracy gains, but instead increases communication overhead and latency.
The research compares five strategies: lexical baseline, dense retrieval + LLM reranking, zero-shot schema-constrained LLM, single-agent Agentic RAG, and a three-agent collaborative team.
Microsoft senior engineer Bob Ward proposed on the Data Exposed show moving chat completion and Agentic RAG directly inside the SQL engine so that AI inference can be completed without data leaving the database.
All Facts (16)
The technical evolution of RAG has gone through three stages: Naive RAG, Advanced RAG, and Modular RAG.
80%VerifiedAgentic RAG grants the model autonomous planning capability, enabling it to assess whether information is sufficient, decide whether to perform multi-round retrieval, choose which tools or data sources to invoke, and carry out self-reflection and iterative correction.
65%VerifiedTraditional RAG follows a unidirectional, fixed, one-shot process, whereas Agentic RAG is a cyclically iterative, multi-step execution process.
65%UnverifiedAgentic RAG用搜索工具替代了固定流水线，让Agent自主决定是否调用搜索、何时重写查询、是否需要多轮检索
90%Unverified传统RAG中大模型只在最后的生成阶段才被使用，而Agentic RAG从用户输入问题的那一刻起就让大模型参与决策
90%UnverifiedAgentic RAG的实现依赖模型具备三类关键能力：规划能力（Planning）、工具调用能力（Tool Use）和多步迭代能力（Iteration）
80%Partially VerifiedThe core idea of Agentic RAG is to encapsulate the various stages of RAG as callable Tools and to endow large models with the ability to make autonomous decisions.
80%UnverifiedAgentic RAG中模型可以在一次用户请求中执行5-10次甚至更多的工具调用
70%UnverifiedAgentic RAG采用'思考→行动→观察→再思考'的循环闭环模式，模型可以在生成最终答案之前进行多步工具调用和动态策略调整
60%UnverifiedIn multi-hop questions, answers are scattered across multiple documents and require finding clue A before knowing to look for clue B; the agent loop transforms retrieval from a one-time action into a repeatedly executable tool call.
50%UnverifiedIn the Google FRAMES benchmark, a simple agent loop (capable of reading retrieved results and searching again) achieved an accuracy of 92.7%.
50%UnverifiedThe three-agent collaborative team doubled system latency without delivering any improvement in accuracy.
50%UnverifiedOn classification tasks with a closed vocabulary, stacking agent roles does not automatically yield accuracy gains, but instead increases communication overhead and latency.
50%UnverifiedArchitectural complexity should be matched to task benefit; when the candidate space is closed and finite, multi-agent negotiation and verification mechanisms may only increase costs without touching the accuracy bottleneck.
50%UnverifiedThe research compares five strategies: lexical baseline, dense retrieval + LLM reranking, zero-shot schema-constrained LLM, single-agent Agentic RAG, and a three-agent collaborative team.
50%UnverifiedMicrosoft senior engineer Bob Ward proposed on the Data Exposed show moving chat completion and Agentic RAG directly inside the SQL engine so that AI inference can be completed without data leaving the database.
50%