Definition
An advanced architecture where an LLM acts as an autonomous agent to orchestrate the retrieval process, using reasoning loops to decide when to search, which tools to use, and how to refine queries based on intermediate results. This approach sacrifices latency and cost for significantly higher accuracy in complex, multi-hop, or ambiguous information retrieval tasks.
Unlike Naive RAG, which follows a static 'retrieve-then-read' script, Agentic RAG uses dynamic loops and self-correction.
"An investigative journalist who follows leads, cross-references sources, and pivots their investigation based on new evidence, rather than a clerk looking up a single file."
- Tool Calling(Prerequisite)
- Naive RAG(Baseline Comparison)
- Query Decomposition(Component)
- ReAct Pattern(Component)
Conceptual Overview
An advanced architecture where an LLM acts as an autonomous agent to orchestrate the retrieval process, using reasoning loops to decide when to search, which tools to use, and how to refine queries based on intermediate results. This approach sacrifices latency and cost for significantly higher accuracy in complex, multi-hop, or ambiguous information retrieval tasks.
Disambiguation
Unlike Naive RAG, which follows a static 'retrieve-then-read' script, Agentic RAG uses dynamic loops and self-correction.
Visual Analog
An investigative journalist who follows leads, cross-references sources, and pivots their investigation based on new evidence, rather than a clerk looking up a single file.