The idea
Some questions need facts scattered across many documents: "Which suppliers are linked to projects that missed deadlines?" Plain chunk retrieval finds pieces but not the connections. GraphRAG extracts entities (people, companies, products) and relationships with an LLM, stores them as a graph, and retrieves by walking the graph.
Microsoft’s GraphRAG also clusters the graph into communities and writes summaries for each, which helps with global questions like "What are the main themes across these 500 reports?"
It is expensive to build (many LLM calls) and harder to keep current. Use it when relationship questions matter; start with normal RAG first.
It answers multi-hop and whole-corpus questions that vector search structurally cannot.
Picture it
Words to know
- Knowledge graphNodes (entities) connected by typed edges (relations).
- Multi-hop questionNeeds combining facts from several places.
Try it without code
Draw the graph
- Read 3 short news articles about one industry.
- Draw people and companies as boxes and relations as arrows.
- Write a question only the drawing can answer.
Use it in code
Extract triples with an LLM and query them with networkx.
import json, anthropic, networkx as nx
client = anthropic.Anthropic()
def triples(text):
msg = client.messages.create(model="claude-haiku-5-5", max_tokens=500, temperature=0,
messages=[{"role": "user", "content":
f"Extract (subject, relation, object) triples as a JSON list of 3-item lists.\n{text}"}])
return json.loads(msg.content[0].text)
G = nx.DiGraph()
for s, rel, o in triples("Acme supplies Project Atlas. Atlas missed its Q3 deadline. Maria leads Atlas."):
G.add_edge(s, o, relation=rel)
print(list(G.edges(data=True)))
print(nx.descendants(G, "Acme"))Build a small graph from 20 documents and answer 3 multi-hop questions by graph traversal plus an LLM summary.
Common mistakes
- Duplicate entities ("Acme", "Acme Ltd") with no merging.
- Building a graph when plain RAG would answer the questions.