Turning Communities into a Global Answer

GraphRAG’s communities become useful only when they can support a response to a particular question. The key focus is the handoff from a reusable hierarchy to a bounded, query-specific answer: first the system builds reports that compress each community, then it maps the question across those reports and reduces the partial answers into a global response. Following that handoff explains both why GraphRAG can cover a corpus-wide question and where ranking, context limits, and compression can still narrow the result.

Community reports are reusable evidence views

A community summary is not just a shorter copy of one document. It is a report about a group of connected entities, relationships, and associated claims. GraphRAG creates these reports before a user asks a question so that the expensive work of organizing the corpus can be reused. A collection that will receive many global questions can pay once to construct a set of views, then use those views as the material for later synthesis.

The summaries are built from the bottom up. At the leaves of the community hierarchy, the system starts with descriptions of graph elements. It prioritizes community edges by the combined degree of their source and target nodes. Degree is a simple measure of how many connections an entity has in the graph, so this rule favors relationships involving prominent elements. The system then adds descriptions of the source node, target node, edge, and related claims to the summary context until the model’s token limit is reached.

This rule gives the leaf report an operating principle: use the available context on relationships that appear structurally prominent, along with the elements that explain them. It is not a claim that the highest-degree relationship is always the most important to a user. A low-degree entity may be central to a narrow question, and an unusual event may matter more than a frequently mentioned one. Degree is a practical prioritization signal for fitting a report, not a complete definition of importance.

At a higher level, the system has two options. If all the element summaries for that community fit within the context window, it can summarize them directly. If they do not fit, it ranks the sub-community summaries by the number of tokens they represent and substitutes shorter sub-community reports for the longer element descriptions associated with them. The parent report can then describe a larger region without placing every underlying element into one prompt.

That substitution is the central compression move. A lower-level report stands in for a group of details when a higher-level report needs to cover a wider scope. The parent retains a description of the sub-community, but it no longer has every underlying description available in full. Recursively, higher communities can incorporate lower-level summaries until the hierarchy provides a broad report that fits the context window.

The result is a family of evidence views rather than one universal summary. A leaf report can preserve a focused relationship and its supporting claims. A parent report can describe how several such regions relate. A root-level report can provide a compact view of the broad graph.