Topics / Knowledge graphs & AI
What is a knowledge graph, explained simply?
The problem as a graph
Here you see a tiny knowledge graph: entities linked by named relations. This is exactly how a machine stores knowledge — not as prose, but as nodes and edges. The empty relation is a connection the system does not yet "know" and would first have to activate (learn). This is a lens on the topic.
Graph as text
- Berlin → Germany (active)
- Germany → Europe (passive)
- Language model (LLM) → Berlin (active)
- Language model (LLM) → Germany (active)
- Language model (LLM) → Verified source (empty)
Step by step
- Pick a piece of knowledge and break it into entities: which "things" appear (places, people, concepts)? These become your nodes.
- Name the relations between them — not just "is related", but concretely: is-capital-of, was-created-by, belongs-to. A named edge is the core of every knowledge graph.
- Mark the state: active = the system is using this relation right now, passive = stored but quiet, empty = never yet learned.
- Find an empty relation as a redirect target — for instance a link to a verified source that an LLM does not yet pull, but could.
- Redirect energy there: instead of letting the model just "guess more", activate the relation to the source (retrieval). Guessed text becomes knowledge you can look up.
- Zoom out: many such node-edge triples form a whole graph — the shared language of search engines, semantic AI and this model.
Seen through the model
Imagine you want to teach a machine: "Berlin is the capital of Germany." A human reads the sentence. A machine prefers to store a triple: node "Berlin", relation "is-capital-of", node "Germany". These three parts — two entities and one named relation — are the smallest piece of a knowledge graph.
Now suppose you add "Germany is-located-in Europe". The system can already infer what you never said directly: Berlin is (indirectly) in Europe. The Germany–Europe relation was passive, stored and quiet; a question activates it. This is how semantic AI "thinks": not in sentences, but in nodes and edges that light up when a question nudges them.
A language model (LLM) like the ones behind modern chatbots works differently at first — it estimates the next word from patterns. But connect it to a knowledge graph or a verified source (retrieval) and you activate the previously empty "LLM–source" relation: guessed text becomes knowledge you can look up. This shows that this site's everyday way of thinking and real AI technology share the same scaffolding — as a shared lens, not as an equation.
Frequently asked
What is a knowledge graph, explained simply?
A knowledge graph is knowledge as a network: entities are nodes (things, places, people), relations are named edges between them (is-capital-of, belongs-to). Instead of storing sentences, a machine stores triples and can infer new connections from them. It is the same node-edge logic this model uses for everyday life.
How do AI systems think in entities and relations?
Semantic AI does not store prose but nodes and named edges. A question acts like a vibration: it activates the matching relations, which then "light up" while the rest stays quiet. Standards like W3C PROV organise the world into Entity, Activity and Agent — a machine-readable version of exactly this entity-relation way of thinking.
How are knowledge graphs and LLMs connected?
An LLM estimates the next word from patterns and may guess. A knowledge graph supplies verified nodes and edges. Connecting both (retrieval) activates a previously empty relation between model and source: probable text becomes knowledge you can look up. One possible reading — a practical way to couple both strengths.
Keep thinking
Related terms: Entity, Relation, Network level