GraphRAG

GraphRAG adds structure to RAG by building a knowledge graph (entities + relations) alongside your vector store. It helps most when questions require multi-hop reasoning, like:

• “Which components depend on X, and what breaks if we change it?”
• “What’s the relationship between A and B across multiple docs?”

Learning goals

• Understand the GraphRAG pipeline (extract → store → expand → retrieve → synthesize)
• Build a toy graph from documents
• Use graph expansion to guide retrieval

The GraphRAG pipeline

1. Entity extraction: find entities (people, concepts, APIs)
2. Relation extraction: connect them (A uses B, A depends on B)
3. Graph store: nodes/edges + attributes
4. Query-time:
   • extract entities from the question
   • expand neighborhood (k hops)
   • retrieve text chunks near those entities
5. Synthesis: answer grounded in retrieved text + graph structure

Toy example (NetworkX)

python

import networkx as nx

G = nx.Graph()
G.add_edge("RAGAS", "RAG Evaluation", relation="measures")
G.add_edge("Hybrid Search", "BM25", relation="uses")
G.add_edge("Hybrid Search", "Embeddings", relation="uses")

def expand(seed: str, hops: int = 1) -> set[str]:
    frontier = {seed}
    seen = {seed}
    for _ in range(hops):
        nxt = set()
        for n in frontier:
            nxt |= set(G.neighbors(n))
        nxt -= seen
        seen |= nxt
        frontier = nxt
    return seen

print(expand("Hybrid Search", hops=1))

How it connects to your vector store

A practical pattern:

• store text chunks in a vector DB
• store entity mentions as metadata on each chunk
• graph expansion picks which entities (and therefore which chunks) to fetch

When GraphRAG is a bad idea

• If your questions are mostly “find the paragraph” (classic RAG is enough)
• If entity extraction is noisy (graph becomes garbage-in-garbage-out)
• If latency budget is tight (graph expansion adds steps)

Mini-lab (optional)

Build a “course knowledge graph”:

• entities: week topics, tools, libraries
• relations: “uses”, “extends”, “evaluated-by”
• query: multi-hop questions and compare to hybrid search