GraphRAG vs. plain RAG: the honest scorecard
A knowledge graph can make AI answers far better, but only for certain questions, and the indexing bill can shock you. Here is what five studies actually found.
Key points
- GraphRAG won 72 to 83 percent of head-to-head comprehensiveness comparisons against standard RAG in Microsoft's own testing.
- Graph-guided retrieval raised average recall scores from 73.4 percent to 87.8 percent on multi-step question benchmarks.
- Building the knowledge graph cost roughly $48 per corpus with GPT-4o, far above a basic vector index.
- On simple, single-fact lookups, plain text chunks matched or beat graph methods in a 2025 Michigan State and Meta study.
- Microsoft's lower-cost follow-up, LazyGraphRAG, cut the indexing price to about 0.1 percent of the original, admitting the full build is too expensive for many teams.
Search the right four words inside a document and you get an instant, useful answer. Ask what the recurring complaints were across two years of customer emails, and the same system shrugs.
That gap is the problem GraphRAG was built to fix.
What is GraphRAG, and why does it exist?
Standard RAG, which stands for retrieval-augmented generation, works like a well-organised filing cabinet. You ask a question, the software finds the most relevant slips of paper (called chunks), and hands them to an AI language model to write an answer. Fast, cheap, good for direct questions.
A filing cabinet can't connect the dots, though. If the answer lives across a dozen files, no individual slip of paper contains it and the model never sees the link.
GraphRAG takes a different approach. Before anyone asks a question, a large language model reads the entire document collection and maps out the named things inside it: people, companies, dates, events, policies, and how those things relate to each other. That diagram is the knowledge graph. The software then groups it into clusters of related topics and writes a plain-English summary for each cluster. When a question arrives, those summaries do the heavy lifting instead of raw document snippets. Our earlier look at enterprise AI agents from 29 July 2026 touched on why this kind of structured, relationship-aware knowledge matters when an AI has to reason across a whole organisation.
When does the graph actually win?
The honest answer, drawn from Microsoft's original research plus four independent studies, is: it depends entirely on the question type.
| Question type | Plain chunks score | Graph score | Winner |
|---|---|---|---|
| Simple fact lookup | 60.9 | 60.1 | Tie |
| Multi-step reasoning | 42.9 | 53.4 | Graph +10 pts |
| Whole-corpus summary | 51.3 | 64.4 | Graph +13 pts |
| Global comprehensiveness | baseline | +72 to 83% win rate | Graph |
For questions that require joining scattered facts or summarising a large body of material, the graph wins clearly. A quick factual lookup? The extra machinery adds cost without adding accuracy.
Recall scores (the share of truly relevant passages the system actually finds) jumped with graph-guided retrieval: from 73.4 percent up to 87.8 percent on average across standard multi-step question tests.
What does it cost, and is the scoring fair?
Two problems deserve plain-language honesty.
Building the graph isn't free. One analysis put the indexing cost at around $48 per corpus using OpenAI's GPT-4o model. Microsoft's cheaper follow-up, LazyGraphRAG, defers most of that work and cuts the price to roughly 0.1 percent of the original. That gap tells you the full build is impractical for many real-world teams.
Many performance comparisons also used another AI model as the judge, which introduces bias. Auditors found that simply swapping the order in which two answers appear can shift the winner by more than 30 percentage points. One method's reported 66.7 percent win rate fell to about 39 percent after corrections.
The large recall gains still hold up. Narrow comprehensiveness margins deserve scepticism.
What should a team actually do?
Route simple fact questions to plain chunks, send complex multi-part questions to the graph. Combining both methods consistently beat either one alone in controlled tests. That's the practical finding, and it's worth taking seriously: there's no universal answer here, only a routing problem.
Pick the right tool for each question, then build the system that makes that choice for you.
