The GraphRAG Index as a Working Model

The GraphRAG index is easiest to understand as a chain that changes the unit of reasoning: a passage becomes extracted elements, repeated elements become graph structure, and graph structure becomes nested communities. The useful judgment is not whether this chain sounds sophisticated, but whether it gives a global question an organized view of the corpus while keeping its uncertainty and compression visible. The model below connects those transformations and identifies the work that still belongs to answer generation.

One piece of meaning, several representations

Consider a source passage that describes an organization working with another organization on a project, along with a date and an outcome. The passage begins as ordinary prose. Its meaning is carried not only by the names, but also by who acted, what the relationship was, when it occurred, and what qualification surrounds the claim. GraphRAG does not carry that entire passage unchanged through every stage. It creates successively more structured representations of it.

First, chunking determines which surrounding words the extraction model sees together. The passage becomes part of a text chunk that is small enough to process. That boundary can preserve useful context or place related statements in separate calls. A longer chunk may reduce the number of calls but can reduce the recall of information that appears early in the chunk. The chunk is therefore the first place where the source’s continuous meaning is translated into a processing unit.

Next, the extraction model proposes semantic elements. It may record the two organizations as entities with types and descriptions, the collaboration as a relationship with a source and target, and the date and outcome as claims. These are not independent facts discovered outside the passage. They are LLM-produced descriptions of what the passage appears to say. If the prompt is tailored to the domain, it may distinguish the kinds of entities and relationships that matter there; if it is too generic, important distinctions may never enter the representation.

The proposal is still an instance, not yet the final graph. The same organization can be mentioned in many chunks, and the same relationship can be extracted repeatedly. Graph construction aggregates those instances into nodes, edges, and associated claims. Entity descriptions are combined and summarized. Relationship observations become an edge, with duplicate counts used as edge weights. The graph now has a reusable structure that can show which entities are connected and which relationships recur in the indexed collection.

Finally, community detection uses that structure to create nested scopes. Connected elements are partitioned into root communities, and communities with further internal structure are partitioned again. A node can therefore have a place in a broad group and in a narrower descendant group.