GraphRAG Community Reports: Turning a Knowledge Graph Into Answers
How GraphRAG builds a graph from text, finds communities with hierarchical Leiden, writes community reports and uses them for global and local search.
This article is published in English.
Sources: Building the Graph with GraphRAG: Community Reports and Summaries
A large pile of documents rarely fails for want of information. It fails because meaning is spread out. One passage names a person, another describes an event and a third explains the outcome, and nothing ties them together.
Conventional RAG retrieves relevant passages well enough, but it tends to stumble on broad or abstract questions, especially those spanning an entire dataset. GraphRAG attacks that gap by constructing a knowledge graph from the text, clustering related material and writing summaries that support reasoning over the whole collection.
A knowledge graph is a network of entities and the links between them. Formally it consists of entities, relationships and triples, each triple joining two things through a specific type of connection. Graphs like this already help with search, fact-checking and question answering, so they suit systems that need more than isolated chunks of text.
From raw text to a graph
Building the graph starts by cutting the source material into smaller units. In GraphRAG's standard pipeline, documents become TextUnits, and the system pulls entities, relationships and sometimes claims out of each one. It then merges duplicate entities and connections and condenses their descriptions into simpler graph objects. The output is a structured index, with tables and embeddings, that preserves the important relationships instead of a flat list of facts.
It sounds simple, but a good graph takes more than spotting names. Research suggests a solid pipeline also has to handle pronouns and aliases, find relationship phrases and tie those phrases to the right entities and actions. In practice that calls for coreference resolution, entity linking, relation linking and triple extraction. If one stage is weak, the graph can be inaccurate before anyone summarizes a thing.
From entities to communities
With the entity graph ready, the next job is finding its structure. GraphRAG groups related nodes into clusters at several levels using hierarchical community detection. The graph stops being a lookup table and becomes a map of themes, subthemes and local topics. Microsoft's documentation describes this as bottom-up summarization for whole-dataset understanding, and Microsoft Research says the hierarchy can run from broad themes down to very specific topics.
The clustering algorithm matters. GraphRAG uses hierarchical Leiden community detection, which has support in network science research. Leiden was designed to guarantee well-connected communities, and a paper in Scientific Reports found it often produces better partitions than Louvain and runs faster. Tighter communities mean better summaries, since the model is condensing text that really fits together.
What a community report holds
Community reports are what make the graph useful. Once communities exist, GraphRAG writes a report for each. The top-level report can describe the entire dataset, while lower-level reports cover smaller groups of entities, relationships and topics. Microsoft says this exposes the structure of the data before a user has asked anything.
A good report is more than a short paragraph. In GraphRAG's standard output it can include a title, a summary, the full content, a relevance rating, an explanation of that rating and a set of roughly five to ten key findings. The documentation says the reports are meant for people to read and for search systems to use, with the short summary serving as a quick version for later steps.
That gives each community two jobs. It helps people grasp what each part of the graph means, and it hands the system reusable context for new questions. The report works as an explanation for humans and as a memory layer for the machine, which is why community reporting isn't an optional extra. It's a main reason GraphRAG can turn raw data into useful answers.
How summaries drive global and local search
Community summaries pay off when a question arrives. The GraphRAG paper argues that regular RAG does poorly on global questions, the ones about an entire dataset instead of one fact. GraphRAG writes the community summaries ahead of time and uses them for these broad queries. The paper's abstract reports substantially more comprehensive and diverse answers to big-picture questions on datasets of around 1 million tokens.
In the official global search flow, reports are processed in batches in a map-reduce style. Each batch produces intermediate points with importance scores, and the highest-scoring points are combined into the final answer. The documentation notes a trade-off: lower levels of the hierarchy give more detailed answers, but cost more time and compute because there are more reports to read.
For narrower questions GraphRAG switches to local search. That mode combines the structured graph with the relevant source text, so the model can answer detailed questions about specific entities with better context. Having both modes shows why community summaries matter. They let the system look at broad themes or zoom into details without rebuilding all the context each time.
Lessons for better community reports
A strong graph begins with careful extraction, not only a good clustering algorithm. When entities are fragmented, aliases go unmatched or relationships are vague, the community report usually sounds generic, since the community itself was poorly defined. Extraction, merging, clustering and reporting work best as a single connected process rather than four unrelated tasks, and that's where the strongest summaries come from.
Operations count too. The GraphRAG repository warns that indexing can be expensive and recommends beginning with a small setup. Version 1.0 is said to ship with simpler configuration, a leaner data model and a smaller storage footprint than before. The guidance also advises tuning prompts, since the defaults won't suit every dataset. In practice, good reports come from iteration: adjust extraction prompts, check cluster quality, revise report prompts and only then scale up.
Another lesson comes from how GraphRAG evolved. Microsoft Research's dynamic community selection adds a filtering step. Before the full global-search pass runs, the system screens the community reports and keeps only those that bear on the question. In a test on AP News data, the method cut the average token cost by 77% at one level while keeping output quality close to the static approach. Searching deeper could add detail, but at higher cost. Takeaway: reports don't all deserve equal processing, so picking the right ones matters about as much as summarizing them well.
What it adds up to
Building the graph is really about creating context. Linking scattered text is the graph's contribution. Community detection sorts that meaning into levels. Reports and summaries turn the levels into something useful to both people and language models. Done well, the summaries aren't decoration. They become the main instrument for making sense of large, messy datasets.
Quick answers
What is a community report in GraphRAG? A text document written by an LLM about one cluster of the graph. Typical fields are a title, a summary, the full body, a rank with its justification and a handful of key findings. People read it, and search steps reuse it later.
Why use community summaries instead of only text chunks? Some questions are global and not local. Both the paper and the documentation argue that broad questions about themes or patterns across a whole corpus are poorly served by plain vector retrieval, while pregenerated summaries give a dataset-wide view that can be recombined into better answers.
Why does Leiden matter? A summary can't be better than the cluster it describes. Leiden yields communities that are well connected, and it scores well in practice on both running time and the quality of the partition. That makes it a sensible foundation for a hierarchy of reports.