GraphRAG uses a knowledge graph during retrieval, often alongside vector search, to expand the context passed to an LLM.
2
Graph structure can improve answer accuracy, support question types that baseline RAG cannot answer, and make retrieved information easier to inspect.
3
Creating the knowledge graph is difficult for unstructured data, but once the graph exists it can simplify application development, debugging, explainability, and governance.
Summary
Emil Eifrem explains GraphRAG through the history of web search. Google improved keyword search with PageRank, then added a knowledge graph so search could work with concepts and relationships rather than text alone. GraphRAG applies the same idea to LLM applications. A vector search first finds relevant documents or nodes. The application then follows graph relationships to retrieve related products, authors, categories, or other context before sending the result to the LLM. Eifrem describes benefits including higher answer accuracy, access to question types that baseline RAG cannot answer, easier debugging, and better explainability and governance. He is direct about the cost: building a graph from unstructured PDFs and text is difficult, while structured sources such as Snowflake, Postgres, MySQL, and Oracle are easier to map into a property graph. He ends with a demonstration of Neo4j's Knowledge Graph Builder, which extracts concepts and relationships from documents, web pages, and YouTube links.
Search moved from matching strings to using relationships between concepts
Eifrem begins with the history of web search. Early companies used keyword search and inverted indexes such as BM25. That worked while the web was small, but the AltaVista effect produced too many results for people to inspect. Google improved this with PageRank, a graph algorithm based on eigenvector centrality. Later, Google's Knowledge Graph added concepts and relationships to the text and links between documents. Eifrem uses this progression to frame GraphRAG as another change in how applications retrieve context.
GraphRAG adds graph retrieval after an initial search
Eifrem defines GraphRAG simply: RAG where the retrieval path uses a knowledge graph. It can still use other methods, including vector search. In his customer-service example, articles about Wi-Fi routers are connected to products, product hierarchies, and the engineers who wrote them. A question about flashing lights and dropped connections starts with vector search. The application then follows graph relationships to find related articles and other context before passing the selected information to the LLM.
The basic retrieval pattern starts with vectors and expands through the graph
The initial vector search gives the application a small set of nodes. Eifrem compares this to an almost-primary-key lookup into the graph. The application walks from those nodes to find related content based on graph structure. If the expansion returns too many nodes, it can rank them and select the top results. PageRank is one possible ranking method. The resulting context goes to the LLM with the original question.
Graph context can improve accuracy and answer additional types of questions
Eifrem cites work from Data.world, LinkedIn, and Microsoft. He says Data.world found response accuracy was three times higher across about 43 questions when a knowledge graph was combined with vector search. He describes LinkedIn's results as showing roughly a 75% or 77% increase in accuracy. Microsoft's work also found that GraphRAG could answer an important class of questions that baseline vector RAG could not answer at all.
GraphRAG becomes easier to develop after the graph has been created
Eifrem says users report that GraphRAG applications are easier to build than baseline RAG applications once a knowledge graph is available. The difficult part is creating that graph in the first place. In an example from a fintech company, moving an application from a vector database to Neo4j improved several operations. The team could calculate embeddings at the database level, retrieve related actions by following relationships, and inspect a visual application cache that helped them find bugs.
Explicit graph data is easier for people to inspect than vector representations
Eifrem compares the sentence 'apples and oranges are both fruit' in vector space and graph space. A graph can show Apple connected to Fruit and Orange connected to Fruit. Vector representations support similarity calculations, but their meaning is opaque to a person inspecting the system. Graph data is explicit, deterministic, and visual. Eifrem says this visibility helped the fintech team fix bugs after moving from a vector-only store to GraphRAG.
Graph structure supports explanation, auditing, and governance
The same explicit relationships that help developers understand an application also help teams operating it. Eifrem connects GraphRAG with explainability, auditability, and governance. When someone needs to explain why a result or decision occurred, the graph provides inspectable data about the entities and relationships involved. He presents this as a production benefit alongside better answers and easier development.
Graph construction is easiest for structured data and hardest for raw documents
Eifrem separates graph inputs into structured data, unstructured data, and data with structured fields containing long-form text. Structured sources such as Snowflake, Postgres, MySQL, and Oracle can be mapped into a property graph model relatively easily. PDFs and raw web text are much harder to convert, and he says the tooling has been immature for a long time. The mixed category is common in enterprise production systems.
Knowledge Graph Builder extracts a graph from documents and links
Eifrem demonstrates Neo4j's Knowledge Graph Builder, which had launched a few weeks earlier. Users can add PDF files, YouTube links, Wikipedia pages, or a cloud storage bucket. His example combines Andrew Ng's newsletter, the OpenAI Wikipedia page, and a YouTube podcast. The tool creates document chunks, embeddings, and logical concept elements such as machine learning, then displays their relationships in a graph. It also includes a chat interface for inspecting returned results.
"This Vector space representation is completely opaque to a human being but the graph representation is very very clear it is explicit it's deterministic it's visual."12:48
Who should watch
You are building a RAG application and vector search is missing relationships between documents, products, people, or categories.
Your team already has structured or semi-structured enterprise data and needs a more inspectable retrieval layer.
You need to debug retrieval decisions or explain why an LLM application returned particular information.
