AI & context
6 min

What Is a Knowledge Graph Database?

Key takeaways

  • Three tiers get called the same thing: graph database, knowledge graph database, and the knowledge graph itself.
  • The middle tier is defined by global identity, standards-based query and declarative constraints.
  • Named graphs matter more than buyers expect. Without them, asserted and inferred statements look identical.
  • It will store and validate your model. It will not decide what a term means or who may change it.
In this article
Share

What is a knowledge graph database?

A knowledge graph database is the storage and query engine underneath a knowledge graph. It holds the nodes, edges and statements, answers queries over them, and in some cases runs inference across them. It is infrastructure, and the phrase describes a middle tier that neither "graph database" nor "knowledge graph" quite captures.

The distinction matters because three things get called the same thing. A graph database stores connected data. A knowledge graph is the modelled, governed meaning. A knowledge graph database sits between them: a store built specifically for the kinds of statements a knowledge graph makes, including identity that survives leaving the database and semantics a machine can reason over.

What are the three tiers, concretely?

Naming them separately is the quickest way to stop talking past each other in an evaluation.

TierWhat it isWhat it does not do
Graph databaseStorage and traversal over nodes and edgesCarry portable identity or machine-readable semantics
Knowledge graph databaseA store built for standards-based statements, with global identifiers and often a reasonerDecide what anything means or who agreed it
Knowledge graphThe modelled, governed meaning itselfStore itself. It needs one of the tiers below

Most market confusion comes from products in tier one describing themselves in tier three language. The knowledge graph vs graph database piece works through that specific collision, and types of graphs maps where each sits.

What should a knowledge graph database actually give you?

Five capabilities, and the last one is the one buyers most often leave off the list.

Capabilities worth checking

  • Global identifiers. A concept identified by a resolvable identifier, so the same entity referenced from three systems is recognisably one thing.
  • Standards-based query. SPARQL over RDF, so queries and data both survive a change of product.
  • Inference, where you want it. Deriving what must also be true from what was stated, with control over when that runs.
  • Declarative constraints. Validation expressed as data, so rules are reviewable by people who are not developers.
  • Named graphs. Separating asserted statements from inferred ones, and one source from another.

The last one is quietly important and often missing. Without it, a statement you asserted and a statement a reasoner derived look identical, and you cannot audit how a conclusion was reached.

Does it have to be RDF?

For a knowledge graph database in the sense above, in practice yes, and the reason is identity rather than ideology.

The properties that distinguish this tier are global identity, standards-based semantics and declarative constraints. RDF, OWL and SHACL were designed to carry exactly those. A property-graph store can be made to approximate them, and the approximation lives in application code. That is the condition that makes a model impossible to govern.

RDF vs property graphs covers that trade in detail. The short version: property graphs are often faster to start with and the identity gap compounds.

How big does this have to be?

Smaller than the market implies, and that changes which products are worth considering.

Vendor benchmarks run to billions of statements, and most enterprise knowledge graphs holding governed meaning are orders of magnitude below that. A vocabulary of ten thousand concepts with their relationships, constraints and provenance is a substantial governed asset and a trivial database workload.

What that implies for a shortlist

  • Scale claims are rarely the deciding factor. Almost every product in the category will handle your volume.
  • Query shape beats query speed. Governed-meaning queries are small, frequent and latency-sensitive, which is a different profile from analytical traversal.
  • Operational burden matters more. A store that needs a specialist to run is a store that gets quietly retired.

The exception is when the same store is also holding operational data at volume. That is a legitimate architecture and it changes the sizing conversation entirely, because you are then buying for the analytical workload and the governed model is a passenger.

What will it not do for you?

Decide anything. This is the part buyers consistently expect and never get.

A knowledge graph database will store your model, validate data against the rules you declare, and answer queries quickly. It will not tell you what "active customer" means, who gets to change that, or which definition applies in which part of the business. Those are governance questions, and no storage engine answers them.

The practical consequence is that buying this tier solves the infrastructure problem and leaves the problem that probably prompted the search. Teams that buy a knowledge graph database expecting the governance to arrive with it spend a year discovering it did not. Ontology management covers the layer that does.

How do you evaluate one?

Mostly on the things that are hard to change later.

What to test

  • Export fidelity. Load a real model, export it, diff. Count what was lost, especially constraints and provenance.
  • Reasoning control. Can you run inference on demand, and tell the derived statements apart afterwards?
  • Constraint validation. Can a domain expert change a rule without a deployment?
  • Scale on your shape. Benchmarks use benchmark data. Load your own and run your own queries.

Performance dominates these conversations and is usually the least decisive factor, because most enterprise knowledge graphs are far smaller than the vendor benchmarks and the queries are shaped differently. The thing that decides the next five years is whether the model can leave. How to evaluate a knowledge graph platform covers the wider process.

So what should I take away?

A knowledge graph database is the middle tier: a store built for standards-based statements with global identity and machine-readable semantics. It is necessary infrastructure and it settles nothing about meaning.

TopQuadrant builds above this tier. Our knowledge graph platform is where the model gets defined with the experts who own it, validated against declared constraints, and made resolvable to the systems that need an answer. The store underneath can be whatever your architecture already runs, provided the model can leave it. Test that with an export and a diff, on your own model.

Give your AI the context it's been missing

See how the TQ Data Foundation turns your enterprise knowledge into trusted, Al-ready context.