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)
Criteria for choosing Agentic RAG: use traditional RAG for simple Q&A and low-latency scenarios; adopt Agentic RAG only for knowledge-intensive, highly variable and complex scenarios.
When engineering an Agentic RAG implementation, a maximum loop round limit must be set to prevent the model from falling into an infinite loop.
For simple scenarios such as Q&A on a company employee handbook, the low latency and low cost of traditional RAG are instead more suitable.
Compared with traditional RAG, Agentic RAG brings higher latency and Token costs.
Agentic RAG in industry has two implementation modes: mode one uses an independent model to make a routing decision on whether retrieval is needed, and mode two registers the knowledge base as a tool for the model to use autonomously through a tool-calling protocol; the latter supports more complex reasoning chains.
Agentic RAG does not force retrieval on every question; instead, it lets the model itself decide whether the current question requires querying the knowledge base, which is closer to real business patterns.
Footnote is an Agentic RAG system that uses an Agent loop for retrieval, returning retrieval results as sentence-level blocks and having the API validate citations, rather than having the model generate citation numbers itself.
In the Google FRAMES benchmark, a simple agent loop (capable of reading retrieved results and searching again) achieved an accuracy of 92.7%.
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.
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.
4 more timeline events
All Facts (20)
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%UnverifiedThe iterative reasoning capability of Agentic RAG introduces engineering challenges such as increased latency, higher costs, and reduced behavioral predictability.
60%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%UnverifiedCriteria for choosing Agentic RAG: use traditional RAG for simple Q&A and low-latency scenarios; adopt Agentic RAG only for knowledge-intensive, highly variable and complex scenarios.
50%UnverifiedAgentic RAG in industry has two implementation modes: mode one uses an independent model to make a routing decision on whether retrieval is needed, and mode two registers the knowledge base as a tool for the model to use autonomously through a tool-calling protocol; the latter supports more complex reasoning chains.
50%UnverifiedCompared with traditional RAG, Agentic RAG brings higher latency and Token costs.
50%UnverifiedFor simple scenarios such as Q&A on a company employee handbook, the low latency and low cost of traditional RAG are instead more suitable.
50%UnverifiedWhen engineering an Agentic RAG implementation, a maximum loop round limit must be set to prevent the model from falling into an infinite loop.
50%UnverifiedAgentic RAG does not force retrieval on every question; instead, it lets the model itself decide whether the current question requires querying the knowledge base, which is closer to real business patterns.
50%UnverifiedFootnote is an Agentic RAG system that uses an Agent loop for retrieval, returning retrieval results as sentence-level blocks and having the API validate citations, rather than having the model generate citation numbers itself.
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%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%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%