Vector vs Graph RAG Performance Gap

In a controlled comparison using a 20-question evaluation on a grant application corpus:

• GraphRAG: 90% correct (18/20)
• Ontology-guided KG with chunks: 90% correct
• Vector RAG baseline: 60% correct (12/20)
• Ontology KG without chunks: 15-20% correct

The gap between vector and graph approaches (60% vs 90%) suggests that for question-answering over domain-specific knowledge, structured retrieval substantially outperforms similarity-based retrieval.

The collapse without chunks (15-20%) shows that structure alone isn’t sufficient, you need the textual content too.

Related: 06-atom—chunk-integration-critical, 07-molecule—vectors-vs-graphs, 07-molecule—rag-core-tradeoffs