GraphRAG: entities, communities and global questions
Questions such as 'what were the main complaint themes?' have no single matching chunk. GraphRAG builds a graph of entities and communities so it can answer them.
Interactive simulation. It needs JavaScript; the text below explains what it shows.
What you are looking at
Step 1 lists the extracted entities, step 2 shows the communities and their summaries, step 3 runs either a global map–reduce over summaries or a local walk from the question's entities. The graph on the right highlights the path. All RAG simulations share one knowledge base of 18 chunks from 11 documents of a fictional outdoor-gear shop, and three questions: one phrased differently from the documents, one that needs two documents, and one about the whole corpus.
Things to try
- Run the complaints question: global search over community summaries.
- Run the fabric question: a local walk from Ridge 2 to Aurora fabric to Lumen Textiles.
- Compare with agentic RAG, which reaches the same facts with search hops.
How it works in real systems
Microsoft's GraphRAG uses an LLM to extract entities and relations, the Leiden algorithm to find communities and LLM-written community reports. Indexing is expensive, so it suits stable corpora with analytical questions.
Code lab
The same idea in real code, step by step. Copy each block into a notebook or a file and run it in order.
Step 1 · Extract entities and relations
An LLM reads every chunk and returns entities and relations as JSON. Build a graph from them.
# pip install networkx openai
import json, networkx as nx
from openai import OpenAI
client = OpenAI() # reads OPENAI_API_KEY
def ask(prompt, model="gpt-4o-mini"):
r = client.chat.completions.create(model=model, messages=[{"role": "user", "content": prompt}], temperature=0)
return r.choices[0].message.content
G = nx.Graph()
for i, c in enumerate(chunks_text := [
"The Ridge 2 tent uses Aurora ripstop fabric.", "Aurora ripstop fabric is woven by Lumen Textiles in Porto.",
"Q3 complaints were mostly about late deliveries.", "Summit 45 zips jammed in cold weather, a Q3 complaint."]):
data = json.loads(ask('Extract entities and relations from the text as JSON {"entities": [..], "relations": [[a, relation, b], ..]}. Text: ' + c))
for e in data["entities"]: G.add_node(e)
for a, r, b in data["relations"]: G.add_edge(a, b, relation=r, chunk=i)
print(G.edges(data=True))
Step 2 · Communities and summaries
from networkx.algorithms.community import louvain_communities # GraphRAG uses Leiden; Louvain is a close cousin
comms = louvain_communities(G, seed=0)
summaries = [ask("Summarise what these facts have in common:\n" + "\n".join(f"{a} {d['relation']} {b}" for a, b, d in G.subgraph(c).edges(data=True))) for c in comms]
Step 3 · Local and global search
Local search walks out from the entities named in the question; global search maps the question over every community summary and reduces the partial answers.
def local(q, hops=2):
start = [n for n in G if n.lower() in q.lower()]
near = set(start)
for _ in range(hops): near |= {m for n in near for m in G.neighbors(n)}
facts = [f"{a} {d['relation']} {b}" for a, b, d in G.subgraph(near).edges(data=True)]
return ask(f"Answer from these facts:\n{chr(10).join(facts)}\nQ: {q}")
def global_search(q):
partial = [ask(f"Using this summary, answer if you can, else say 'none': {q}\n{s}") for s in summaries]
return ask(f"Combine these partial answers into one: {partial}\nQ: {q}")
Microsoft's implementation does all of this from the command line: graphrag init --root ./kb, graphrag index --root ./kb, graphrag query --root ./kb --method global --query "...".