GraphRAG earns its cost on global questions, multi-hop chains, and entity-rich corpora: the queries plain top-k vector search cannot stitch together. It is overkill for single-fact lookups and never chosen for speed.
Imagine a huge pile of news clippings. If someone asks 'what does this one article say,' you just find that clipping and read it. That is normal search. But if someone asks 'across all these clippings, what are the big recurring storylines, and how are these two people connected through a chain of events,' flipping through clippings one at a time will not work. So beforehand you build a giant relationship map (who knows whom, what links to what) and a set of summaries grouping related clippings. That map is GraphRAG. It costs effort to build, but it answers the big-picture and connect the dots questions a plain search never could. For a simple 'what is the refund policy' question, the map is wasted work.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
GraphRAG arrived as the answer to a question plain vector RAG simply cannot handle: "summarize the main themes across this entire corpus." Ask top-k retrieval that, and it returns five chunks that happen to be similar to the words "main themes". That is useless, because the answer is a property of the whole corpus, not of any chunk you can retrieve.
What matters is the boundary. GraphRAG is powerful and expensive. Its cost is justified in a narrow band of cases, while in a much larger band plain retrieval is the correct, cheaper choice.
Global and multi-hop queries break local retrieval for a structural reason. Everyday lookups belong on vector RAG, and choosing GraphRAG "for speed" gets it backwards.
What GraphRAG builds before any query arrives
The defining trait of GraphRAG is that it front-loads work into an offline indexing pipeline. Plain vector RAG does almost nothing offline beyond chunking and embedding; all the intelligence is the query-time similarity search.
GraphRAG does much more. It runs an LLM over the corpus to extract entities (people, organizations, concepts) and the relationships between them, assembling a knowledge graph. It then runs community detection, typically a Leiden-style algorithm, to find clusters of densely connected entities. Finally it summarizes each community, often hierarchically, so there are summaries at multiple levels of granularity.
That output (a graph plus a tree of community summaries) is the asset. It encodes structure that lives across documents, not inside any one of them.
The cost is real. Entity and relationship extraction is an LLM call over the whole corpus, so indexing a large corpus can cost real money and hours. This is the trade at the heart of GraphRAG: pay heavily once, offline, to answer questions you otherwise could not answer at all.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- Microsoft's GraphRAG extracts an entity-relationship graph, runs Leiden community detection, and summarizes communities to answer global sensemaking questions over a corpus.
- Analysts querying 'what are the recurring risk themes across these thousands of incident reports' use community summaries, which top-k chunk retrieval cannot assemble.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow does GraphRAG actually answer a global question that no single chunk contains?
It runs community detection over the entity graph, summarizes each community, then map-reduces partial answers from those summaries into a global response.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Picking GraphRAG to cut latency: graph construction and traversal usually make it slower and more expensive to build, not faster; it is chosen for the questions it can answer, never for speed.
60 second bullets to scan on the way to the call.
What GraphRAG builds offline: entity graph plus community summaries
Why global/aggregative questions need community summaries, not chunks
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.