chapter eleven

11 Correction, planning, and reasoning

 

This chapter covers

  • Why retrieval quality alone is not sufficient for RAG systems
  • How control, reasoning order, and validation determine real-world RAG performance
  • How and why RAG architectures are shifting from single-pass to agentic
  • When to use corrective control (CRAG) versus planning-first retrieval (LevelRAG)
  • How retrieval failures map to architectural decisions

In chapter 10, we explored RAPTOR and optimizing the index by organizing information into a semantic tree to address the pervasive issues of Lost in the Middle and Incorrect Specificity (FP6). RAPTOR RAG showed how a hierarchical index can improve RAG quality.

Even with better indexing, standard Retrieve-then-Generate pipelines share a structural weakness: they are single-pass. The system bets on a linear, optimistic chain (vector search finds the right document, the document contains the answer, the LLM extracts it correctly) and ships whatever comes out the other end.

11.1 The agentic turn

11.1.1 The enterprise compliance assistant

11.1.2 Anatomy of a single-pass failure

11.1.3 Stale knowledge and missing freshness checks

11.1.4 Retrieval collapse on long-tail entities

11.1.5 Multi-hop reasoning gaps

11.2 Corrective retrieval augmented generation

11.2.1 The retrieval evaluator

11.2.2 Decompose-then-recompose

11.2.3 Corrective actions and web search integration

11.2.4 The CRAG pipeline

11.2.5 Evaluation, shortcomings, and where CRAG fits

11.3 LevelRAG

11.3.1 The high-level searcher

11.3.2 The low-level searcher

11.3.3 LevelRAG implementation

11.4 Reactive versus proactive agency

11.5 Agentic RAG in practice

11.5.1 The agentic spectrum

11.5.2 Products and platforms

11.5.3 How CRAG and LevelRAG fit together in production

11.5.4 Refusal as a feature

11.6 Summary