Concept Contrast Fable

Retrieval-augmented generation and graph retrieval-augmented generation

Conclusion first:RAG finds material in a document library and feeds it to a large model; Graph RAG finds material along entity relationships in a graph database. The latter is a variant of the former—the underlying retrieval switches from a pure text index to a graph structure, but the overall process remains within the RAG framework.

RAG finds material in a document library and feeds it to a large model; Graph RAG finds material along entity relationships in a graph database. The latter is a variant of the former—the underlying retrieval switches from a pure text index to a graph structure, but the overall process remains within the RAG framework.

Fable

Asking the way by flipping through books vs. asking the way by following a map

The old city library had a reference desk. Readers came with questions, and the librarian's workflow had not changed in decades: first break down the keywords in the question, go to the shelves to find relevant titles, pull out the corresponding passages, excerpt and organize them into an answer, and hand it to the reader. Everyone called this process "asking directions by flipping through books."

One year, a cartographer came to the library and spent more than half a year extracting the people, events, organizations, and places from the library's documents and drawing them into a huge relationship map, which was hung on the wall by the reference desk. Lines connected the figures on the map: “Sima Qian” was connected to “Records of the Grand Historian,” “Emperor Wu of Han,” and “Grand Astrologer”; “Grand Astrologer” was in turn connected to “astronomical observation,” “calendar revision,” and “court records.”

At first the consultant only occasionally looked up at this map to help clarify his thinking. Later he developed a new way of asking: when a reader asked, "Who influenced Sima Qian when he wrote the Records of the Grand Historian?", he did not just search the shelves for the keywords "Sima Qian," "Records of the Grand Historian," and "influence." Instead, he first looked on the map for Sima Qian's adjacent nodes—teachers, predecessors, contemporaries, authors he cited—and then followed each relationship line on the map to retrieve the corresponding documents from the shelves in a targeted way. This new way of asking was called "asking the way by map."

The difference between the two ways of asking for directions isn't in the result format, but in the retrieval path. Asking by flipping through books follows a direct line from keyword to shelf. Asking by following the map adds an intermediate stop: first look at the map, find a relationship subgraph from it, then use each edge of the subgraph to precisely locate documents on the shelf. The map's contribution is that retrieval no longer guesses by keywords, but follows known relationships.

But asking directions by map has a precondition: the map must already be built. Without that map, asking directions by map falls back to asking directions by flipping through books. And maintaining the map is an ongoing cost — every time the library's collection is updated, the map must update a batch of nodes and edges. The cartographer cannot draw it just once.

The help desk later layered the two ways of asking for directions: every question first starts with keywords and goes through the book-flipping process, while at the same time looking for the corresponding entity node on the graph. If there is a relevant subgraph on the graph, the graph-following process is added on top, using the documents guided by the subgraph as semantic reinforcement. If the graph is empty or no corresponding node can be found, the graph-following process silently skips and falls back to the pure book-flipping path.

At an industry sharing session, the curator said one thing: turning pages to ask the way is a basic skill, and following the map to ask the way is a bonus. The bonus must be built on top of the basic skill, not the other way around.

Later, a sign was hung at the entrance of the consultation desk: Window One handles page-turning, Window Two handles page-turning plus image lookup. Readers who came first went to Window One, and Window One only redirected them to Window Two when it found that relationship clues were needed. The two windows shared the same row of bookshelves.

Retrieval-Augmented Generation (RAG)

A general technical paradigm that retrieves relevant text fragments from a document library and inputs them as context into a large model to enhance generation quality. The retrieval layer is usually a vector or keyword index and does not depend on graph structure.

Graph-Enhanced Retrieval Generation (Graph RAG)

A variant of RAG that replaces the retrieval layer with a knowledge graph: locate relevant entities and relationship subgraphs on the graph, extract associated text fragments along edges, then hand them to the large model for generation. The retrieval path uses the graph's structural information, not just text similarity.

Story Mapping

Flip through books to ask the way (keywords → bookshelf → excerpts → answer): the standard RAG process, text retrieval plus generation.
Relationship graph (connecting people, events, organizations, and places): knowledge graph.
Follow the map to ask for directions (view the map first → follow relationship lines → retrieve targeted literature → generate): the Graph RAG process.
If the graph is empty, fall back to flipping through books: Graph RAG depends on a pre-built knowledge graph, otherwise it degrades to standard RAG.
The layering of Window One and Window Two: RAG is the baseline, Graph RAG is the incremental overlay.

In the project

Before doing Graph RAG, first confirm that basic RAG is already working and has clear shortcomings—usually scenarios where standard RAG recall is insufficient, such as multi-hop reasoning and entity relationship queries. Do not start by building a knowledge graph. Graph RAG adds a layer of ongoing maintenance cost: entity extraction, relationship updates, and version synchronization for the graph. In addition, Graph RAG and standard RAG should share one generation layer and diverge only in the retrieval path. This way you can easily run A/B tests: for the same question, compare the answer quality difference between pure RAG and RAG plus Graph RAG.

Retrieval-Augmented Generation Graph-enhanced retrieval generation