-
Notifications
You must be signed in to change notification settings - Fork 2
Software Ontologies
An ontology is a formal description of the concepts, relationships, and rules that define a domain, written in a format that software can read. When tools and agents can read that description directly it can improve their understanding of a domain. By using the defined processes or events for the objects in the ontology the AI can understand the verbs of the domain and thus understand exactly how objects can be modified via processes or events.
Most software systems share data through APIs and databases but do not share meaning. Two systems can exchange the word "customer" and mean different things. One tracks billing accounts, and the other tracks registered users. Ontologies exist to solve that problem.
An ontology is a formal, explicit specification of the concepts, properties, and relationships that exist within a domain. The term comes from philosophy, where it means the study of what exists. In software it is purely practical. It is a structured vocabulary that makes knowledge machine-readable and supports reasoning.
The key ingredients of any ontology are:
-
Classes: the types of things in the domain, e.g.,
Product,Customer,Transaction - Properties: the attributes of those things. Product has a price, a SKU, and a category.
- Relationships: how things connect to each other. Customer places Order, and Order contains Product.
- Axioms: logical rules and constraints that must hold. An Order must have at least one Product.
- Instances: the actual data. These are specific customers, products, and orders that conform to the schema.
The dominant standard for expressing ontologies is OWL (Web Ontology Language), built on top of RDF (Resource Description Framework). Together they let you describe a domain so that reasoning engines can look up existing facts and infer new ones.
These three are often conflated but serve distinct roles:
- A database schema defines how data is stored. It is implementation-specific and optimised for query performance.
- An ontology defines what things mean and how they relate. It is implementation-neutral and optimised for reasoning and interoperability.
- A knowledge graph is a populated ontology. The ontology is the schema, and the graph is the data instantiated within it.
An ontology can unify multiple databases under a shared semantic model without moving any data.
Ontologies are not a theoretical exercise. Some of the most important software systems in the world are built on them:
The Systematized Nomenclature of Medicine, Clinical Terms, is an ontology with over 350,000 medical concepts used in electronic health records across more than 50 countries. A clinical system in the UK and one in Australia can exchange a patient record and know that "myocardial infarction" means exactly the same thing in both.
The Financial Industry Business Ontology, maintained by the EDM Council, provides a shared semantic model for financial instruments, legal entities, and transactions. Regulators and banks use it to ensure that a "derivative" or a "counterparty" means the same thing across institutions and jurisdictions.
The collaborative vocabulary backed by Google, Microsoft, and Yahoo defines a lightweight ontology for web content. When a search engine reads a page and understands that a block of text is a Recipe with a cookTime and a list of ingredients, Schema.org makes that possible.
Microsoft's Fabric platform now ships a first-class Ontology item that defines business entities, relationships, rules, and actions across an organisation. AI agents and Real-Time Intelligence components consume it as a shared context layer. Every agent reasons from the same definition of Revenue, Customer, or Product.
Google's AudioSet project defines a hierarchical ontology of over 630 audio event classes, from Speech and Music down to Finger snapping and Thunderstorm. Researchers use it to build one of the largest labelled audio datasets in the world, and Google uses it to organise and classify knowledge across its research systems.
Amazon's Alexa team published research on the Alexa Meaning Representation Language (AMRL), a semantic framework that maps natural language utterances to structured meaning representations grounded in an ontology of intents, entities, and relationships. It is a direct example of ontology-driven NLU. The ontology defines what concepts Alexa can understand, and the meaning representation language bridges human speech to those concepts.
The most important word in modern AI is grounding. A large language model is a probabilistic text predictor. Without external constraints, it will generate answers that sound plausible but are factually wrong. The industry term for this is hallucination, and it is the primary barrier to deploying AI reliably in important domains.
Ontologies address hallucination directly:
When a retrieval-augmented system (RAG) is backed by a knowledge graph, it can only assert relationships that actually exist in the graph. If your ontology does not contain a link between Product X and Feature Y, the agent cannot invent one. This reduces hallucinations compared to plain vector search.
A vector database retrieves text that is similar to a query. An ontology-backed system can infer new facts. It can answer "which customers are at risk?" by traversing relationships the user never explicitly asked about. The result is closer to a reasoning engine than a search engine.
In a multi-agent architecture, each agent needs a consistent understanding of domain terms. Without a shared ontology, Agent A's "account" and Agent B's "account" may refer to entirely different things. A shared ontology keeps agents aligned on what each term means.
Schemas change. Business rules change. An ontology separates what things mean from how they are stored, so when your database migrates, the meaning layer stays intact.
Regulated industries need to prove that a decision was made correctly. An ontology makes the reasoning chain explicit and inspectable. Neural networks alone cannot satisfy this requirement because their reasoning is not inspectable.
Getting started with ontologies does not require building from scratch:
Protégé (Stanford University) is the leading open-source ontology editor. It supports the full OWL 2 specification, integrates with reasoning engines like HermiT and Pellet, and runs both as a desktop application and in the browser via WebProtégé. Most academic and enterprise ontologists use it first.
Microsoft Fabric Ontology brings ontology-driven AI to organisations already on the Microsoft stack. It auto-generates an initial ontology from existing Power BI semantic models and exposes it to Copilot agents and Real-Time Intelligence workflows through a structured API.
The Ontology Playground on GitHub provides examples and patterns for building and querying ontologies. Teams can use it to explore ontology-oriented development in practice.
FIBO, SNOMED CT, and Schema.org are all freely available as starting points for domain-specific work. Reusing an established ontology is almost always faster and more interoperable than building one from scratch.
If you are new to ontologies, the fastest way to get started is:
- Identify one bounded domain, e.g., a product catalogue, a customer model, or an incident taxonomy. Keep it narrow at first.
- Define the core classes and relationships in plain language before touching any tooling. This is the hardest and most valuable step.
- Model it in Protégé using OWL. Run a reasoner to check for contradictions and infer implicit relationships.
- Populate it with real data from an existing database or API. This is when a schema becomes a knowledge graph.
- Connect it to an AI layer, e.g., a RAG pipeline, a Fabric IQ agent, or a custom LLM integration. Observe the improvement in response quality and consistency.
The investment is upfront and real. An ontology is one of the few artifacts in a software project that grows more valuable over time. Every new system you connect to it benefits from the shared semantic layer.
Ontology research dates to the early 1990s at Stanford and Xerox PARC. The W3C standardised RDF in 1999 and OWL in 2004. For two decades ontologies were primarily an academic and enterprise data management concern. The arrival of LLMs and agentic AI has made them urgently practical. The single biggest limitation of generative AI in enterprise settings is the absence of a shared, machine-readable understanding of what a business's data means. Ontologies provide that understanding.