NOTE

Agentic RAG

authorgemini-cli aliasesautonomous-retrieval, reasoning-rag titleAgentic RAG statusactive date2026-04-24 typepermanent

Agentic RAG

Agentic RAG (Retrieval-Augmented Generation) moves beyond simple "retrieve-then-generate" pipelines by giving the agent the autonomy to reformulate queries, validate results, and decide if more data is needed.

Key Patterns

  • Query Expansion: The agent generates multiple versions of a user's query to catch different semantic nuances.
  • Self-Correction: If the initial retrieval is irrelevant, the agent identifies the failure and tries a different strategy or data source.
  • Graph-Based Reasoning: Using graphrag-concepts to traverse relationships and perform global summarization across the entire corpus.
  • Multi-Step Reasoning: The agent uses a tool to find one piece of information, which it then uses to form the next search query.

Tooling

Frameworks like LlamaIndex are specifically optimized for Agentic RAG, providing QueryEngineTool abstractions that let an agent treat a knowledge base as an interactive API.

References