Why this page exists. Ontology work fails far more often from starting too big than
from choosing the wrong tool. This page covers the components, the formalisms, the roles, the stores
and the deployment choice — but the section that matters most is the last one:
how to start without boiling the ocean.
1 · Ontology vs taxonomy vs schema
2 · Where you actually need one
3 · What it contains
4 · Shape & form
5 · Who builds it
6 · Where it lives
7 · PaaS or on-prem
8 · Standards to adopt
9 · Ontology as code
10 · How to start
1 · Ontology vs taxonomy vs schema concept
These five get used interchangeably and cost real money when confused in a design review.
| Artefact | What it captures | Example | Answers |
|---|---|---|---|
| Glossary | Business terms and their agreed definitions | "Active Customer means…" | What does this word mean? |
| Taxonomy | Hierarchical classification — is-a only | Product → Vaccine → mRNA Vaccine | What kind of thing is this? |
| Ontology | Entities, relationships, constraints and rules that support inference | Study investigates Compound; Compound treats Indication | How do things relate, and what follows from that? |
| Knowledge graph | An ontology populated with instances | Study NCT0123 investigates GSK-456 | What is actually true in our estate? |
| Data model / schema | Physical structure of a table or document | dim_product(sku, name…) | How is it stored? |
| Semantic layer | Business metrics and dimensions over physical tables | revenue = sum(net_amount) | How do we calculate it consistently? |
The one-line distinction: a taxonomy tells you what something is; an ontology
tells you what it is connected to and what can be inferred. If your requirement never
needs a relationship traversal or an inference, you need a taxonomy or a glossary — and you have just
saved yourself a year.
2 · Where you actually need one trade-off
| Use case | What the ontology buys you | Could you skip it? |
|---|---|---|
| Impact & blast-radius analysis | "What breaks if I change this table, service or field" — transitive, across systems | No — this is inherently multi-hop |
| Entity resolution / master data | The same customer, site or compound reconciled across a dozen systems | No — identity is a relationship problem |
| Text-to-SQL and NL query | Which tables mean what, and how they legitimately join | Partly — a semantic layer may be enough |
| Agent tool selection & context | Which system owns which concept, so an agent picks the right tool | Often — a good tool catalogue may suffice |
| Hybrid graph + vector RAG | Multi-hop questions vector search cannot answer | Usually yes at first — start with hybrid search |
| Policy & classification propagation | Tag a concept once, inherit the controls everywhere it appears | No — inheritance is the point |
| Regulatory interoperability | Speaking IDMP, CDISC or FHIR because a regulator requires it | No — the standard is the ontology |
| Merging systems after acquisition | A common conceptual layer across incompatible models | No — this is the classic case |
The qualifying test: write down five real questions the business cannot answer today. If none
of them requires following a relationship through another entity, you do not need an ontology
yet. Most "we need an ontology" programmes are actually asking for a glossary, a semantic layer, or
better search.
3 · What it contains concept
| Component | What it is | Skipping it costs you |
|---|---|---|
| Classes / entities | The things that exist — Study, Site, Compound, Incident, Service | Nothing to hang anything on |
| Properties | Attributes with datatypes and units | Ambiguity — is duration in hours or days? |
| Relationships | Typed edges with domain and range — investigates, depends-on, supersedes | This is the ontology. Without it you have a taxonomy |
| Hierarchies | Subclass and subproperty — what inherits from what | Every rule has to be restated at every level |
| Cardinality & constraints | A Study has exactly one Sponsor; a Site has at least one Investigator | Data that validates but is nonsense |
| Axioms & rules | What can be inferred — transitivity, symmetry, disjointness | You infer in application code instead, inconsistently |
| Identifiers & namespaces | Stable URIs or keys that survive renaming | Nothing can be merged or federated later |
| Labels & synonyms | Preferred term, alternates, abbreviations, other languages | Search and NL query miss the way people actually speak |
| Mappings / crosswalks | Links to external standards and to physical columns | An ontology disconnected from data — the classic shelfware failure |
| Provenance & versioning | Who asserted this, when, from what source, under which version | No way to audit or roll back a meaning change |
Mappings are the component teams skip and the one that decides success. An ontology
with no link to physical columns is a diagram. The mapping layer — concept → table → column — is what
turns it into something a query engine or an agent can actually use.
4 · Shape & form trade-off
| Formalism | Looks like | Strengths | Costs | Use when |
|---|---|---|---|---|
| RDF / OWL | Turtle, JSON-LD; SPARQL; SHACL | W3C standard, formal inference, federates, aligns with public ontologies | Steep learning curve; scarce skills; verbose | Regulated, interoperability, external standards |
| SKOS | Concept schemes, broader/narrower | Lightweight, ideal for vocabularies and taxonomies | Not expressive enough for real constraints | Controlled vocabularies, thesauri |
| Labeled property graph | Neo4j + Cypher | Pragmatic, fast, easy to hire for, great tooling | No standard semantics; inference is yours to write | Operational graphs, impact analysis, agent memory |
| LinkML | YAML schema | Author once, generate OWL, JSON Schema, SHACL, Python, SQL DDL | Younger ecosystem | You want one source and several outputs |
| JSON Schema / Pydantic | Code-facing types | What applications and LLM tool definitions actually consume | No inference, no relationships beyond nesting | The operational projection — generate it, do not hand-write it |
| Semantic layer | dbt / Cube / MetricFlow YAML | Consistent metrics; already in the analytics workflow | Metrics and dimensions only — not a general ontology | BI consistency and text-to-SQL grounding |
You will need two representations, and that is fine. One authoritative
(OWL, SKOS or LinkML) and one operational (property graph, JSON Schema, semantic layer). The
rule is that the second is generated from the first. The moment both are hand-maintained they
diverge, and the ontology becomes documentation nobody trusts.
5 · Who builds it concept
| Role | Owns | Failure if they are absent |
|---|---|---|
| Domain SME / steward | The meaning. Which distinctions matter and which are noise | The commonest failure. A technically elegant ontology the business does not recognise |
| Ontologist / knowledge engineer | The formalism — classes, relations, constraints, alignment to standards | Inconsistent modelling that cannot be reasoned over |
| Data engineer | Mappings to physical tables, population pipelines, refresh | An ontology unconnected to real data |
| Architect | Scope, boundaries, which standard to adopt, what stays out | Scope creep to the whole enterprise; nothing ships |
| Governance / QA | Change control, versioning, regulatory alignment | Meaning drifting silently; no audit trail |
| LLM assistance | Proposes candidate entities, relations and synonyms from documents and schemas | Slow manual extraction — but it only ever proposes; a human ratifies |
A workable modern split: an LLM pass over existing schemas, documents and tickets
proposes candidate concepts and relationships; the ontologist shapes them; the SME ratifies meaning;
CI validates the result. That collapses the slowest part of the work without handing meaning to a
model — propose automatically, ratify deliberately.
6 · Where it lives landscape
Four distinct storage questions, and teams routinely conflate them.
Authoring & source of truthGit. The ontology is code — versioned, reviewed, diffable, released. Never a database as the master copy.
Instance storeThe populated knowledge graph — triple store or property graph.
Metadata & governanceCatalog holding classification, lineage and policy — Unity Catalog, Purview, Collibra.
Operational projectionGenerated JSON Schema, semantic-layer YAML, Cypher constraints that applications consume.
| Option | Category | Strengths | Choose when |
|---|---|---|---|
| Authoring & validation | |||
| Protégé / WebProtégéOSS | Authoring | The de-facto OWL editor; reasoners built in | Formal OWL work — the default |
| LinkMLOSS | Authoring | YAML source, generates OWL, SHACL, JSON Schema, Python, DDL | You want ontology-as-code with real outputs |
| SHACL / pySHACL, ShExOSS | Validation | Constraint validation you can run in CI | Always — this is your test suite |
| VocBench, Neo4j ArrowsOSS | Authoring | Collaborative vocabulary editing; quick model sketching | SME-facing curation and whiteboarding |
| TopBraid EDG, PoolParty, Semaphore, metaphactorycommercial | Authoring | Governance workflow, SME-friendly UIs, publishing | Large curated vocabularies with non-technical stewards |
| Triple stores — RDF / SPARQL | |||
| Apache Jena FusekiOSS | Store | Free, standard, easy to start | Modest scale, getting going |
| GraphDB (Ontotext)free tier | Store | Strong reasoning, mature tooling | Inference matters |
| Stardog, AllegroGraph, RDFox, Virtuosocommercial | Store | Virtual graphs over relational data, high-performance reasoning | Enterprise scale with federation needs |
| Oxigraph, BlazegraphOSS | Store | Lightweight, embeddable | Small footprint or embedded use |
| Property graphs | |||
| Neo4jOSS + SaaS | Store | Most mature; Cypher, tooling, vector index; AuraDB managed | Default for an operational graph |
| Memgraph, FalkorDB, KùzuOSS | Store | In-memory speed; FalkorDB for request-path latency; Kùzu embedded | Latency-sensitive or per-tenant graphs |
| Neptune (AWS), Spanner Graph (GCP), Cosmos DB Gremlin (Azure)cloud | Store | Managed, IAM-integrated; Neptune speaks both RDF and property graph | Cloud-standardised estates |
| Apache AGE (Postgres)OSS | Store | Graph queries in the Postgres you already run | Modest graphs, one fewer system |
| Catalog & governance | |||
| Unity CatalogDatabricks | Catalog | Lineage, tags, classification, permissions on real data | Databricks estates — put mappings here |
| Microsoft PurviewAzure | Catalog | Estate-wide classification and glossary | Azure estates |
| DataHub, OpenMetadata, Apache Atlas, EgeriaOSS | Catalog | Open metadata models, glossary, lineage | Multi-cloud or vendor-neutral requirement |
| Collibra, Alation, Atlancommercial | Catalog | Stewardship workflow for non-technical users | Formal governance programmes |
| Semantic layer | |||
| dbt Semantic Layer, CubeOSS + SaaS | Semantics | Metric definitions consumed by BI and text-to-SQL | Consistent metrics is the actual requirement |
| Databricks Metric Views, AtScale, Power BI modelsmixed | Semantics | Native to the platform | Already standardised there |
7 · PaaS or on-prem trade-off
| Factor | Points to managed / PaaS | Points to self-hosted / on-prem |
|---|---|---|
| Data residency & regulation | Region-pinned managed service meets the requirement | GxP validated environment, or data that may not leave the datacentre |
| Content sensitivity | Ontology schema is rarely sensitive — often fine anywhere | Instances may be highly sensitive — patient, trial, commercial |
| Scale | Elastic growth, unpredictable load | Known, stable volumes |
| Latency | Batch or interactive analytics | Graph reads on a request path next to the app |
| Ops capacity | No team to run a triple store | Existing DBA and platform capability |
| Validation burden | Vendor-qualified service reduces your evidence load | You control the change window and the qualification |
| Exit cost | Standards-based (RDF/SPARQL) keeps it portable | Full control, but you carry upgrades and patching |
The split that usually wins in regulated estates: keep the ontology itself in Git
— it is text, it is reviewable, it moves anywhere, and it is rarely sensitive. Then choose the
instance store on the sensitivity of the data, not of the model. Schema in the open,
instances where policy requires. That separation also makes the deployment decision reversible.
8 · Adopt before you build important
In life sciences and healthcare the ontologies largely already exist, are maintained by people whose job that is, and are often what a regulator expects to receive. Building your own version of MedDRA is an expensive way to be non-standard.
| Domain | Established standards | Where to find them |
|---|---|---|
| Clinical terms | SNOMED CT, ICD-10/11, LOINC (labs), RxNorm, UMLS | BioPortal, UMLS Metathesaurus |
| Adverse events | MedDRA | MSSO (licensed) |
| Products & substances | IDMP (ISO 11615/11616), GINAS, ChEBI | ISO, EMA, EBI |
| Clinical trials data | CDISC — SDTM, ADaM, CDASH, Controlled Terminology | CDISC library |
| Real-world / observational | OMOP Common Data Model | OHDSI |
| Interoperability | HL7 FHIR | HL7 |
| Biology & disease | Gene Ontology, UniProt, EFO, MONDO, NCIt | OLS (EBI), BioPortal |
| General / linked data | schema.org, SKOS, Dublin Core, PROV-O, DCAT | W3C, schema.org |
The rule: adopt the standard, then extend it with a small local namespace for what is
genuinely yours — internal programme codes, site hierarchies, system names. Your ontology should be
mostly imports and a thin layer of local concepts. If it is mostly local, either you are in an unusual
domain or you have not looked hard enough.
9 · Ontology as code pattern
# repo layout — the ontology is a versioned, tested, released artefact ontology/ imports/ # pinned versions of external standards core/ # your entities and relationships (.ttl or LinkML .yaml) mappings/ # concept -> physical table.column shapes/ # SHACL constraints = the test suite generated/ # NEVER hand-edited # CI — the same shape as any other build - validate: pyshacl -s shapes/ -d core/ # constraints hold - reason: check no unsatisfiable classes # logically coherent - diff: report added/removed/renamed concepts vs the last release - generate: linkml → JSON Schema, OWL, Python, Cypher constraints - publish: versioned artefact + changelog # semver: renames are breaking
Treat a concept rename as a breaking change and version accordingly. Downstream
queries, mappings and prompts refer to concept names; changing one quietly is the ontology equivalent
of renaming a public API column.
10 · How to start — and how it fails practice
- Start with five questions the business cannot answer today. The ontology exists to answer them; anything not needed for those five waits.
- Adopt the standards that already cover your domain. Import, do not re-model.
- Ten to twenty core entities. Enough for the five questions, no more.
- Write the mappings immediately — concept to real column. An unmapped concept is a hypothesis.
- Populate one real slice end to end and answer question one. A working thin slice beats a complete model.
- Put it in Git with SHACL in CI from day one, not after it grows.
- Generate every downstream form. Hand-maintained copies diverge within weeks.
- Let a model propose, a human ratify. LLM extraction over schemas and documents removes the drudgery without outsourcing meaning.
| Failure mode | What it looks like |
|---|---|
| Boiling the ocean | Eighteen months modelling the enterprise; nothing queryable; funding withdrawn |
| No consumer | Built by a central team with no application depending on it. Nothing keeps it honest |
| Formal purity over usefulness | Beautiful OWL axioms; no mappings; no data; no answers |
| Unmapped ontology | Concepts that never connect to a physical column — a diagram, not infrastructure |
| No versioning | Meanings drift, downstream queries break, nobody can say when or why |
| Rebuilding a standard | A bespoke adverse-event hierarchy that no regulator or partner recognises |
| Two hand-maintained copies | The OWL and the property graph disagree; both are now untrusted |
The one measure worth tracking: how many production queries, applications or agents
depend on the ontology. If that number is zero six months in, the problem is not the modelling — it is
that nothing needed it. A single dependent consumer does more for quality than any amount of review.
Related: AI Engineering Blueprint — where
the ontology sits among the other 36 concerns ·
Graph & memory frameworks — GraphRAG, Graphiti and the
graph databases underneath.