Agentic RAG

Summary

Agentic RAG is the evolution of traditional RAG (Retrieval Augmented Generation) that adds agent capabilities — autonomous decision-making through “think → act → observe” loops — replacing the fixed single-pass retrieval pipeline with iterative, adaptive information gathering.

Key Points

Traditional RAG follows a fixed pipeline: query → retrieve → generate. It lacks task decomposition, adaptive retrieval, and multi-hop reasoning
Agentic RAG makes the LLM a controller that dynamically decides what to retrieve, when, and how — transforming passive retrieval into active decision-making
Two implementation pathways:

1. Tool-Driven (Prompt Engineering + Tools)

Exemplified by Chatbox (36.8k GitHub stars):

query_knowledge_base — semantic search for candidates
get_files_meta — metadata for strategic decisions
read_file_chunks — precision reading of specific segments
list_files — browse file inventory

Key insight: “给模型配备合适的工具和策略性的 Prompt，就能展现出令人惊叹的智能” — appropriate tools + strategic prompts yield remarkable intelligence.

2. RL-Driven (Reinforcement Learning)

Exemplified by Search-R1:

Model learns when/what to search through policy optimization
Enables “推理-搜索-推理” (reasoning-search-reasoning) cycles
Higher adaptability but significantly more complex to implement

Comparison

	Traditional RAG	Tool-Driven Agentic	RL-Driven Agentic
Decision	Fixed pipeline	Rule-based	Learning-optimized
Retrieval	Single pass	Multiple passes	Adaptive multi-pass
Adaptability	Low	Medium	High
Complexity	Low	Medium	High

Relationship to Other Concepts

Directly implements the Augmented LLM and Orchestrator-Workers patterns from Agentic Patterns
The tool-driven approach aligns with Tool Use as Meta-Ability — tools enable deterministic, reliable operations
RAG is increasingly a foundational component within Coding Agents, not a standalone system

Open Questions

When is RL-driven Agentic RAG worth the implementation complexity vs. tool-driven?
How does Agentic RAG interact with prompt caching — do iterative retrieval loops break cache efficiency?
What’s the optimal tool granularity for knowledge base access?

Evidence Timeline

2026-04-08: Initial compilation from Chaofa Yuan’s “RAG 进化之路” (published 2025-10-03, modified 2026-03-18)

My Brain Wiki

探索