What is a knowledge graph?

Many questions that matter in an organisation are about connections. Who owns this service? Which customers use the product affected by this incident? Who approved access to this system, and who do they report to? In tables, each answer means writing joins across several sources. In a graph, it means walking from one node along named edges, which is easier to express and easier to explain.

Knowledge graphs have also become a way to ground language models. Neo4j describes GraphRAG as grounding models with a knowledge graph to improve accuracy and explainability. A vector search finds passages that sound similar to a question; a graph can return the exact relationship the question asks about, with a path that shows how the answer was reached.

There are two common models. The W3C's RDF standard expresses every fact as a triple of subject, predicate and object, where the subject and predicate are identifiers (IRIs) and the object is an identifier or a literal value such as a date. A set of triples is an RDF graph, and shared identifiers let graphs from different sources merge. The property graph model, used by Neo4j, stores nodes and relationships that each carry a label and their own properties.

Either way, the graph needs an organising principle. Neo4j's term covers anything from a simple list of node and relationship types to a formal ontology. The schema stops one source calling an employee a user and another calling them a person without anyone noticing the two are the same.

Building one is mostly integration work: extract entities and relationships from source systems, resolve duplicates (the same person appears in the directory, the HR system and the ticket tool), and record where each fact came from. Queries then use a graph query language to walk from node to node. The hard part is not storage but keeping facts current and knowing which ones to trust.

A security engineer needs to know who may approve a request for access to the billing database. The graph holds four facts drawn from three sources: the billing database is owned by the payments team (from the service catalogue), Priya is a member of the payments team (from the directory), the payments team lead role is held by Priya (from HR), and team leads approve access to their own systems (from policy).

A single query walks from the database to its owning team, from the team to its lead, and returns Priya with the path that justifies it. If the HR record is a month old and the directory shows Priya moved teams last week, a well-built graph shows the conflict instead of picking one. That is why provenance and freshness on each edge matter as much as the edges themselves.

The Swfte company brain, a customer-hosted appliance, includes a time-aware graph of the organisation. Today it syncs people, groups, reporting lines, organisational units and accounts from directories such as Active Directory, Microsoft Entra ID, Okta and Google Workspace, read-only. History is insert-only, so you can read the graph as it stood at an earlier time, and every fact carries one of seven evidence statuses: observed, corroborated, verified, inferred, stale, disputed or unknown.

Its scope is narrower than the word graph may suggest. Collectors for code, cloud, tickets and business systems are on the roadmap, as are the graph explorer and ownership maps. The local API already serves the data. Cortex, Studio and Nexus do not read from the brain yet; that wiring is also on the roadmap.

• The company brain graph
• Sources the brain reads
• Enterprise graph

Definitions on this page were read on the sources below on 7 October 2026. Where sources define the term differently, the page says so. The full glossary lists more terms.