GraphRAG vs vector RAG: retrieval tradeoffs explained

Cloud Codes compares vector retrieval with knowledge graphs for questions that require facts from several documents. The speaker's central argument is that similarity search suits straightforward lookups, while graph traversal can help connect scattered information.

The discussion covers Microsoft's GraphRAG and LazyGraphRAG. The latter moves summarization to query time; the speaker cites indexing costs at 0.1% of full GraphRAG and query costs more than 700 times lower than its global search in a 5,500-article test. A separate $50 indexing estimate uses a different corpus from the cited $33,000 example, so these figures are not a direct comparison.

Reported GraphRAG-Bench results favor graphs for complex reasoning, 53.4 versus 42.9, but show little difference for simple fact retrieval. HippoRAG combines a knowledge graph with Personalized PageRank. KGGen addresses entity resolution, though its higher fact recovery did not produce a clear advantage in the cited question-answering test.

Anthropic's example assigns extraction to Haiku and entity resolution to Sonnet. The speaker also explains why the coding assistant Claude Code reportedly dropped its vector database: code already has structure that ordinary search can navigate. Extraction errors and judge bias limit the benchmark claims. The practical recommendation is to choose retrieval around the questions and the structure already present in the source material.