chapter twelve
12 Graph-based RAG
This chapter covers
- Why vector search fails at multi-hop, relationship-heavy questions
- Graph-Based RAG taxonomy: global summarizers, associative reasoners, and hybrid optimizers
- Hierarchical community detection (Leiden algorithm) and Map-Reduce summarization in Microsoft's GraphRAG
- How Personalized PageRank (PPR) lets HippoRAG mimic the human hippocampus and associative memory
- Dual-level retrieval in LightRAG that routes queries between graph structures and vector indices
- Measuring whether the graph earned its cost, using interaction information to detect passages that only answer the question together
RAG has evolved by improving the resolution of similarity. We moved from keyword matching to dense embeddings, and we got better at finding chunks that look like the question. The chunks themselves, though, still sit in the index as isolated points: close to other chunks that share vocabulary, but otherwise unconnected. The pipeline measures distance, not relationship.