Models & technology

Knowledge Graph

A knowledge graph stores entities such as people, places, and products together with the relationships between them. It is used in AI search and in RAG.

Information stored as connections

A knowledge graph stores information about people, places, companies, products, and events together with the relationships between them.

Connect person A works at company B with company B developed product C, and a path exists between person A and product C. Rather than listing information in tables, it represents it as a network of nodes and edges.

How AI uses it

  • Distinguishing which person or product a search term refers to
  • Exploring relationships between people, departments, documents, and projects
  • Recommending based on connections between products or works
  • Investigating chains in fraud or dependency analysis
  • Supplying relationship-aware material to a model in RAG

In AI search it distinguishes different entities sharing a name and follows several relationships from one question. In Graph RAG, retrieval can follow connections between people, organizations, and events rather than only text similarity.

How it differs from a vector database

A vector database converts text or images into numbers and excels at finding things with similar meaning. A knowledge graph excels at following explicitly stated relationships — who made what, which component goes into which product.

They are complementary rather than competing. Finding relevant text through vector search and confirming relationships through a graph gives a model more precise context.

Cautions

A knowledge graph assumes the underlying information is correctly organized. Register stale information or a wrong relationship and the model will answer from it.

Who defines the relationships, and how updates continue, matters as much. A path existing in the graph does not prove causation or fact. For consequential decisions, check the underlying data and its source.

Related terms

Sources and review information

Last reviewed July 20, 2026

Back to the AI Glossary

Search this site