Stop graphing everything: When GraphRAG actually beats vector RAG
If you have built anything with retrieval-augmented generation (RAG) in the last two years, you have lived its central frustration: You chop your documents into chunks, embed them, retrieve the top few that look similar to the question, and hand them to the model. For “What was our Q3 refund policy?” This works beautifully. For “What are the recurring themes across two years of customer…
Retrieval-augmented generation (RAG) systems have struggled with generating comprehensive answers for complex, multi-hop questions. The issue lies in the fact that isolated chunks of text are insufficient for connecting seemingly disparate pieces of information that form the foundation of a nuanced response. Recent research has introduced GraphRAG, a technique that builds a knowledge graph of entities and relationships within the corpus before feeding it to the model.
This approach aims to improve the ability of RAG systems to answer global, multi-hop questions by leveraging the structure provided by the graph.
The key findings from four independent benchmark studies suggest that GraphRAG performs substantially better than standard vector RAG in specific types of queries. Firstly, in global sense-making tasks where the model needs to understand the entire corpus, GraphRAG outperformed vector RAG in comprehensiveness and diversity, achieving 72-83% comprehensiveness and 62-82% diversity improvements.
Secondly, on multi-hop retrieval benchmarks such as MuSiQue, HotpotQA, and 2WikiMultiHopQA, graph-guided retrieval significantly increased Recall@5, jumping from 73.4% to 87.8%. This improvement is particularly notable in the hardest cross-document sets, with gains of 31 points on MuSiQue and 28 points on 2Wiki.
However, the benefits of GraphRAG are not universal. A 2025 study compared RAG against four GraphRAG variants under uniform conditions, revealing that the effectiveness of GraphRAG depends on the nature of the question. For simple fact retrieval, both methods performed similarly, with no clear advantage. On the other hand, complex reasoning tasks saw a clear advantage for GraphRAG, with scores of 53.4% compared to 42.9% for text chunks.
Contextual summarization also favored GraphRAG, achieving a 64.4% accuracy rate against the 51.3% of text chunks.
Despite these advantages, there are notable drawbacks to implementing GraphRAG. The process of constructing the knowledge graph is computationally expensive, requiring significant resources, and can cost around $48 using GPT-4o for a moderate-sized corpus. Additionally, the quality of the answers generated by GraphRAG is ultimately contingent on the performance of the underlying language model.
While GraphRAG can dramatically improve the quality of answers to complex questions, it is not a one-size-fits-all solution and should be carefully evaluated based on the specific requirements of the application.
Written by urgent.news from VentureBeat's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
