Agentic RAG

Agentic RAG embeds autonomous AI agents into the retrieval-augmented generation pipeline, enabling dynamic decisions about when, what, and how to retrieve based on task requirements.

Traditional RAG follows a fixed flow: query → retrieve → generate. Agentic RAG adds a reasoning layer that can:

• Decide when to retrieve: not every query needs external knowledge
• Reformulate queries: rewrite based on initial results or task decomposition
• Iterate multi-step: retrieve, reason, retrieve again as needed
• Coordinate tools: use retrieval alongside other capabilities (calculation, API calls, etc.)

Two paradigms have emerged:

• Predefined reasoning: structured pipelines with fixed decision points (rule-based when to retrieve, what to do with results)
• Agentic reasoning: the model autonomously orchestrates tool interaction during inference

This shift parallels the evolution from static websites to interactive applications, from passive content delivery to dynamic, context-aware behavior.

The term reflects broader AI agent architectures: reflection, planning, tool use, and multi-agent collaboration applied specifically to the retrieval-generation problem.

Related: 05-atom—rag-definition, 05-molecule—rag-evolution-timeline