HypothesisVery high confidenceCurrent
A query router combining sparse, dense, temporal and graph retrieval is better supported by current evidence than using GraphRAG for all enterprise questions.
Benchmark by query class.
Valid from2026-08-23
Last reviewed2026-08-23
Review again2026-11-21
Dataheart stanceexploring
Evidence for
The systematic 2025 comparison found conventional RAG stronger on single-hop detail and GraphRAG stronger on multi-hop and diverse summarization.
RAG vs. GraphRAG: A Systematic Evaluation and Key Insights · arxiv.org ↗From RAG to Memory: HippoRAG 2 · arxiv.org ↗current project repository · github.com ↗HippoRAG 2 combines graph propagation with passage retrieval and online model use.
RAG vs. GraphRAG: A Systematic Evaluation and Key Insights · arxiv.org ↗From RAG to Memory: HippoRAG 2 · arxiv.org ↗current project repository · github.com ↗Microsoft GraphRAG indexing is expensive and its reference implementation is in maintenance mode.
RAG vs. GraphRAG: A Systematic Evaluation and Key Insights · arxiv.org ↗From RAG to Memory: HippoRAG 2 · arxiv.org ↗current project repository · github.com ↗Evidence against
Results vary by dataset, model, evaluation metric and graph construction.
RAG vs. GraphRAG: A Systematic Evaluation and Key Insights · arxiv.org ↗From RAG to Memory: HippoRAG 2 · arxiv.org ↗current project repository · github.com ↗A high-quality curated enterprise graph may outperform document-derived graphs in additional workloads.
RAG vs. GraphRAG: A Systematic Evaluation and Key Insights · arxiv.org ↗From RAG to Memory: HippoRAG 2 · arxiv.org ↗current project repository · github.com ↗How this claim can fall
Replace this placeholder during human review with an observable result that would materially weaken or falsify the claim.
A published claim remains useful only while it remains possible to prove it wrong.