Knowledge graph

A knowledge graph represents information as a network of entities, people, products, companies, concepts, connected by labeled relationships, such as "acquired by" or "belongs to category," rather than as rows in a table. This structure makes relationships between things first-class and queryable, not something that has to be reconstructed with joins across separate tables.

It differs from a relational database, which is organized around fixed table schemas, and from embeddings-based semantic search, which captures similarity of meaning but not explicit, labeled relationships; a knowledge graph can answer what connects one entity to another precisely, while a vector search can only suggest what looks related. Some systems combine both, using a knowledge graph for structured facts and embeddings for fuzzy matching, an approach sometimes used to ground retrieval-augmented generation more reliably than text retrieval alone.

Knowledge graphs matter for domains with rich, well-defined relationships, fraud rings, supply chains, organizational structures, and for powering question-answering that needs multi-hop reasoning across connected facts. They are closely related to a data catalog or business glossary effort, since both require agreeing on entity definitions. The main pitfall is the cost of building and maintaining one: a knowledge graph decays quickly if the entities and relationships are not kept up to date as the underlying business changes.