====== Agentic RAG ======
Agentic Retrieval-Augmented Generation (Agentic RAG) is an advanced paradigm that integrates autonomous AI agents into the retrieval-augmented generation pipeline. Unlike basic RAG, which follows a static retrieve-then-generate pattern, Agentic RAG employs agents capable of autonomous decision-making, iterative refinement, and dynamic workflow orchestration to handle complex, multi-step information needs.


<mermaid>
graph TD
    Q[User Query] --> P[Plan Retrieval Strategy]
    P --> R[Retrieve Documents]
    R --> E{Evaluate Sufficiency}
    E -->|Insufficient| RF[Reformulate Query]
    RF --> R
    E -->|Sufficient| S[Synthesize Answer]
    S --> ANS[Final Answer]
</mermaid>

===== Background =====
Traditional RAG systems retrieve documents from a knowledge base and concatenate them into the LLM context for generation. This single-pass approach fails on tasks requiring multi-hop reasoning, adaptive query strategies, or cross-source synthesis. Agentic RAG addresses these limitations by embedding agent capabilities directly into the retrieval loop.

The key insight is that retrieval should be treated as an //agentic action// rather than a static preprocessing step. The agent decides //when// to retrieve, //what// to retrieve, and //how// to reformulate queries based on intermediate reasoning. Even sophisticated implementations with 20+ subtasks maintain orchestrated, sequenced, and bounded operations rather than unconstrained exploration, representing a structured workflow approach applied to RAG systems.(([[https://cobusgreyling.substack.com/p/the-ai-agent-reality-gap|Cobus Greyling (LLMs) - Agentic RAG Pipeline (2026]]))

===== Architectural Patterns =====
Singh et al. (2025) propose a comprehensive taxonomy of Agentic RAG architectures organized around four core agentic design patterns:((Singh et al. "Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG." [[https://arxiv.org/abs/2501.09136|arXiv:2501.09136]], 2025.))

  * **Reflection**, Agents evaluate their own retrieval results and generation quality, triggering re-retrieval when confidence is low
  * **Planning**, Autonomous decomposition of complex queries into retrieval sub-plans with dependency tracking
  * **Tool Use**, Integration of heterogeneous retrieval tools (dense retrieval, sparse search, SQL, APIs) selected dynamically per sub-query
  * **Multi-Agent Collaboration**, Distributed retrieval tasks across specialized agents with coordination protocols

The taxonomy classifies systems into: single-agent RAG, multi-agent RAG, hierarchical RAG, corrective RAG, adaptive RAG, and graph-based RAG.(("Agentic RAG Frameworks, Extended Analysis." [[https://arxiv.org/abs/2506.10408|arXiv:2506.10408]]))

===== Autonomous Retrieval Planning =====
In Agentic RAG, the agent autonomously plans [[retrieval_strategies|retrieval strategies]] based on query complexity. Given a user query $q$, the agent generates a retrieval plan $P = \{(q_1, s_1), (q_2, s_2), \ldots, (q_n, s_n)\}$ where each $q_i$ is a sub-query and $s_i$ is the selected retrieval source.

The planning process uses a reward signal $R(P, q)$ that estimates plan quality:

$$R(P, q) = \sum_{i=1}^{n} \alpha_i \cdot \text{rel}(q_i, q) \cdot \text{cov}(D_i, q)$$

where $\text{rel}(q_i, q)$ measures sub-query relevance to the original query, $\text{cov}(D_i, q)$ measures coverage of retrieved documents $D_i$, and $\alpha_i$ are learned weighting coefficients.

===== Adaptive Query Reformulation =====
Unlike static query expansion, Agentic RAG reformulates queries iteratively using feedback from previous retrieval rounds. The agent maintains a belief state $b_t$ at each step $t$:

$$b_{t+1} = \text{Update}(b_t, D_t, \text{Eval}(D_t, q))$$

If the evaluation function $\text{Eval}(D_t, q)$ indicates insufficient coverage or relevance, the agent generates a reformulated query $q_{t+1}$ conditioned on both the original query and the gap analysis.

===== Code Example =====
<code python>
from [[langchain|langchain]].agents import AgentExecutor, create_react_agent
from [[langchain|langchain]].tools import Tool
from langchain_core.prompts import PromptTemplate

def build_agentic_rag(llm, retriever, tools):
    """Build an Agentic RAG pipeline with autonomous retrieval planning."""
    retrieval_tool = Tool(
        name="knowledge_retrieval",
        func=retriever.invoke,
        description="Retrieve relevant documents from the knowledge base"
    )
    all_tools = retrieval_tool + tools

    prompt = PromptTemplate.from_template(
        "You are a retrieval agent. Analyze the query complexity, "
        "plan retrieval steps, and iteratively refine until sufficient "
        "evidence is gathered.\n\nQuery: {input}\n{agent_scratchpad}"
    )

    agent = create_react_agent(llm, all_tools, prompt)
    return AgentExecutor(agent=agent, tools=all_tools, max_iterations=10)
</code>

===== Applications =====
Agentic RAG has demonstrated strong results across domains:((AgenticRAG-Survey GitHub Repository. [[https://github.com/asinghcsu/AgenticRAG-Survey|github.com/asinghcsu/AgenticRAG-Survey]]))
  * **Healthcare**, Multi-hop diagnostic reasoning linking symptoms to disease pathways
  * **Finance**, Adaptive risk modeling with iterative evidence gathering from market data
  * **Legal**, Cross-reference analysis across statutes, case law, and regulatory documents
  * **Education**, Personalized tutoring with dynamic knowledge retrieval (([[https://arxiv.org/abs/2501.09136|Singh et al. (2025), Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG]])) (([[https://github.com/asinghcsu/AgenticRAG-Survey|AgenticRAG-Survey GitHub Repository]])) (([[https://arxiv.org/abs/2506.10408|Agentic RAG Frameworks, Extended Analysis]]))

===== See Also =====
  * [[rag_in_ai|Retrieval-Augmented Generation (RAG) in AI]]
  * [[rag_vs_mcp|Differences Between RAG and MCP]]
  * [[rag_phases|Phases of a RAG System]]
  * [[hermes_orchestration_vs_context_rag|Hermes Multi-Agent Orchestration vs Context+RAG]]
  * [[rag_wrapper_layers|RAG Wrapper Layers]]

===== References =====