Knowledge Graphs (Neo4j)

Entities as nodes, relations as typed edges; query with Cypher, power GraphRAG

RAG & Retrieval · Intermediate · ~5 min

In plain English

Facts stored as connections — this person works at that company, which owns this product — so you can walk the links to an answer instead of searching text.

Why it's worth your time

It's the structure behind entity-heavy domains, and the thing that makes multi-hop questions answerable exactly rather than approximately.

If you remember three things

Nodes are entities, edges are relationships
Queries traverse; they don't rank
Entity resolution is the hard part, not the storage

Overview

A knowledge graph stores information as entities (nodes) connected by typed, directed relationships (edges), with key-value properties on both — a property graph. Neo4j is the leading native graph database for this, queried in Cypher, whose MATCH clause draws the pattern you want like ASCII-art of the graph. Because edges are first-class, multi-hop questions are cheap traversals, and the same structure powers GraphRAG, retrieving a connected subgraph as grounded context for an LLM.

In an interview

A knowledge graph models data as nodes for entities and typed, directed edges for their relationships, with properties on both — a property graph. It's stored in a native graph database like Neo4j and queried with Cypher, where you MATCH a pattern instead of writing JOINs. Multi-hop questions — friend-of-a-friend, supply-chain paths — are cheap traversals. For RAG this enables GraphRAG: retrieve a relevant subgraph as structured, grounded context, in contrast to vector search, which fetches semantically similar text but knows nothing about how facts relate.

Production defaults

Model first: decide your entity and relationship types before ingesting anything. Retrofitting a schema is painful
Resolution: budget real effort for deduplicating entities. It's where these projects succeed or fail
Combine: graph for exact relationships, vectors for fuzzy recall. Use each for what it's good at

What breaks

Traversals return nothing — Entities that should be one node are several. Resolve duplicates before blaming the query.
The graph is stale — No update path from the source systems. A graph that isn't maintained decays faster than an index.

Watch it explained

What is a Knowledge Graph? — IBM Technology, 5:36