Careeroza

GraphRAG Overview

Retrieval over Relationships

Some questions need facts that are connected, not similar: "Which service centre handles the brand of the item in my order?". The answer needs order → item → brand → service centre. GraphRAG extracts entities and relations from documents into a knowledge graph and walks the graph to collect connected facts.

Example

Example

A tiny knowledge graph of triples and a two-hop walk. The collected facts are what you would put in the prompt.

TRIPLES = [
    ("SE1042", "contains", "Boat Airdopes 141"),
    ("Boat Airdopes 141", "brand", "boAt"),
    ("boAt", "service_centre", "boAt Care, Koramangala, Bengaluru"),
    ("boAt", "warranty", "1 year"),
    ("SE1077", "contains", "Prestige Svachh 3L"),
]
graph = {}
for s, rel, o in TRIPLES:
    graph.setdefault(s, []).append((rel, o))

def walk(start, hops):
    facts, frontier = [], [start]
    for _ in range(hops):
        nxt = []
        for node in frontier:
            for rel, o in graph.get(node, []):
                facts.append(f"{node} --{rel}--> {o}")
                nxt.append(o)
        frontier = nxt
    return facts

for f in walk("SE1042", 3):
    print(f)

Output

SE1042 --contains--> Boat Airdopes 141
Boat Airdopes 141 --brand--> boAt
boAt --service_centre--> boAt Care, Koramangala, Bengaluru
boAt --warranty--> 1 year

When It Is Worth It

Use GraphRAG whenStay with vector + hybrid RAG when
questions need multi-hop links (people, products, suppliers)answers live in one or two passages
you need "global" summaries across many documentsdocuments are FAQ or policy style
you already have structured data (a product catalogue, an org chart)budget is tight: building the graph costs many LLM calls

Tools: Microsoft's open-source GraphRAG, Neo4j with vector indexes, LlamaIndex's property graph index. Often a plain SQL join over your own tables is the best "graph".