When to use graphs in RAG
Much use and abuse of words, marketing claims, fluff and fake statements. Sometimes subtle, sometimes explicit. AI generated websites, AI generated articles, dashboards and stats. For some reason we went from ‘data science’ to ‘AI’ and completely dropped ‘science’. Claims and analysis don’t need evidence anymore.
It ain’t my style to shoot at anyone, but a widely-shared graph RAG production write-up does something rare: it contradicts itself in public, inside a single article 😣 In the opening we go from “accuracy jumped from 43% to 91%”, halfway we read “…real production numbers 95% GraphRAG vs. 40–60% plain RAG…” and we end the report with “91% vs. 58%”. Pick a baseline, any baseline, no dataset is named, no task or methodology, no metric. The piece also reports indexing cost at $7.13, and cost-per-document at $0.0048. Decimal-point precision on numbers that come from nowhere reproducible. It’s precision theater, circus progress.
Knowledge graphs and AI context architecture has a citation problem that hides behind good sounding details. Of course, none of this means that (graph) RAG doesn’t work. If you do need solid references, have a look at the following:
- When to use graphs in RAG: https://arxiv.org/abs/2506.05690
- GraphRAG bench: https://arxiv.org/abs/2506.02404
- RAG vs GraphRAG: https://arxiv.org/html/2502.11371v3