Standards
14 min

Taxonomy and Ontology Platforms Compared: How the Categories Differ

Key takeaways

  • Products in this market adopt each other's vocabulary, so a shortlist often compares things built for different jobs.
  • Six categories separate them, from RDF databases to metrics semantic layers.
  • Decide the category before you compare features. A grid laid across categories measures nothing.
  • Every claim here is sourced to the vendor's own published documentation, dated.
In this article
Share

A team evaluating semantic platforms usually ends up with four or five products on a shortlist that look interchangeable from the outside. All of them talk about knowledge graphs. Most mention RDF, SPARQL and ontologies. Several say "semantic layer". The demos all show a graph.

The resemblance is not accidental. Every vendor in this space has an incentive to appear in every shortlist, and the easiest way to do that is to adopt the vocabulary of categories you are not actually built for. The result is a shortlist where the products answer different questions, and a feature grid laid across them rewards whichever product the grid was written around.

The useful question is not which product is better. It is which category your problem belongs to, and then which product in that category has actually built the thing instead of labelling it.

Six categories of platform are routinely compared as though they were one. They differ in what they treat as the asset: the triples, the graph, the query surface, the agreed meaning, the catalogue entry, or the metric definition.

Why do these platforms look alike and behave differently?

Because they overlap at the edges and differ at the centre. An RDF database can hold an ontology. A taxonomy manager can expose a SPARQL endpoint. A data catalogue has a business glossary that looks like a vocabulary. Each of those is a real capability, and in each case it sits at the edge of what the product was built to do, away from its core.

What separates the categories is the asset each one is designed to protect. That decides where the engineering went, which capabilities are first class, and which are a page in the documentation with nothing behind it.

It is worth being blunt about the direction the borrowing runs. Databases describe themselves in the language of meaning far more often than meaning-management products describe themselves in the language of storage. A product whose core asset is the triples has every reason to claim the semantic ground, because the semantic ground is where the budget is.

What are the six categories?

CategoryWhat it treats as the assetRepresentative products
RDF database / triplestoreThe triples. Storage, SPARQL, reasoning at scale.GraphDB, Virtuoso, Amazon Neptune
Property graph databaseThe graph. Traversal performance on a non RDF model.Neo4j, TigerGraph
Graph database with federation and light modelingThe query surface over data left where it lives. Modeling tools exist to serve the query.Stardog, Anzo
Taxonomy and ontology managementThe agreement itself, and the record of who made it.TQ Data Foundation, PoolParty, Semaphore
Data catalogue / metadata managementThe catalogue entry. Discovery, lineage, policy over existing assets.Collibra, Alation, Atlan
Metrics semantic layerThe metric definition, for BI and analytics consistency.dbt, Cube, AtScale

What does each category actually do?

Six categories, taken one at a time. For each, what it is built to answer and where it stops.

RDF database and triplestore

Stores RDF triples, answers SPARQL queries and runs inference over them. Graphwise describes GraphDB, formerly sold under the Ontotext brand, as a "Semantic Graph Database" compliant with RDF 1.1 and SPARQL 1.1. That description is accurate, and it is worth taking at face value. Reading "semantic" as a governance claim would be a mistake. The asset is the data. Editorial control over meaning is not the job, and the documentation shows it.

Property graph database

Graph storage and traversal on a model that is not RDF. Fast and widely adopted, with the trade that the model is the vendor's own, with no W3C standard behind it, so portability between products is not something you get for free.

Graph database with federation and light modeling

This is the category that argues hardest about what it should be called, and the naming is the whole dispute. Products here are graph databases that can query data in place, and not only data loaded into them. Federation is the real engineering achievement and it is a substantial one.

They also ship modeling tools, and the depth of those tools is where the category gets oversold. A modeling capability built to shape a query is a different artifact from one built to hold an agreement. The first needs to describe data accurately enough to retrieve it. The second needs to survive disagreement, record who settled it, and still be answerable about that decision three years later. Products in this category build the first and describe it with the vocabulary of the second.

Stardog positions itself as "The Knowledge Graph-Powered Semantic Layer" and a "semantic AI platform" that "unifies enterprise data where it lives", and since June 2026 as a "Semantic Control Plane". Four layers of semantic vocabulary over what the documentation describes as a database with a federation layer and a model editor. Unifying access to data where it lives is hard and Stardog does it well. It is not the same problem as establishing what the data means and keeping that agreement stable while the business argues about it. A federated query surface will serve two contradictory definitions of "active customer" from two source systems quickly, in one result set, and tell you nothing about which one was agreed.

Taxonomy and ontology management

Authoring, governance, versioning and review of agreed meaning. Progress Semaphore's Knowledge Model Management module describes its job as to "develop, manage and maintain models (thesaurus, taxonomy, ontology or other controlled vocabulary)". PoolParty, part of Graphwise, ships a four state approval workflow moving concepts through Draft, Suggested, Approved and Deprecated.

The asset is not the data and not the query. It is the agreement, and the evidence of who agreed it.

Data catalogue and metadata management

Finds assets, traces lineage and applies policy. The business glossary in a catalogue is genuinely useful, and it is a different thing from a taxonomy management system. It describes assets that already exist. It does not govern the vocabulary itself as a versioned, approvable artifact.

Metrics semantic layer

Holds metric definitions so that revenue means the same thing in every dashboard. This is the term that has drifted most. "Semantic layer" originally described the model layer and now most often means the metrics layer. If a vendor says semantic layer, make them say which one they mean, because the two sit at opposite ends of the stack.

What does governance look like outside category four?

This is where the shortlist separates, and it separates further than most evaluations discover before signing.

Take a bank classifying financial instruments. The classification binds risk reporting, client suitability and regulatory submissions. A change to one definition is a change to all three, so someone has to propose it, someone else has to approve it, and in eighteen months an auditor will ask who did both.

What the documentation shows when you ask for that

  • Roles that do not carry rights. GraphDB offers custom roles, and its own documentation states that they "have no inherent rights or restrictions associated with them, and are instead just markers that can be freely granted to any number of users". A marker labels a user. It does not control an action. Access control in name.
  • Constraints you must commit to before you have a model. GraphDB supports SHACL validation, with the condition that "SHACL support in a given repository must be enabled when that repository is created. You cannot modify an already existing repository by enabling the validation afterwards." Governance is a decision you make on day one, before you know what you are modelling, and it is irreversible in either direction.
  • Control sold separately. GraphDB's fine grained access control "requires a GraphDB Enterprise license". PoolParty's approval workflow "is only available from PoolParty Enterprise Server level" and is "available as an additional module for PoolParty Advanced Server". In both cases the capability that makes the product governable is not in the product you were quoted.
  • An approval process with a hole in it. PoolParty's workflow is the real thing in this group, and its own documentation records the limit: "Deletion of concepts are not covered by an approval process at the moment. This means a deleted concept can not be rejected to bring back a deleted concept." An approved taxonomy can lose a concept with no review and no way to reverse the loss.
  • Change detection in place of change history. Stardog's schema versioning computes "a 64-bit hash from the contents of the reasoning schema graphs" so that it "can be checked to determine if any of the associated named graphs have been modified". It tells you that something moved. It does not tell you what, or who. Stardog is treated at length below.

None of this is a criticism of these products at what they are for. GraphDB is a capable database and PoolParty has the most developed editorial workflow of the three. It is a criticism of buying them to solve a governance problem, which is what the shared vocabulary encourages.

The Semantic Control Plane, read against its own documentation

Stardog deserves separate treatment, because it makes the strongest governance claim in the category. In June 2026 it announced a new brand identity advancing what its own press release calls "the concept of the Semantic Control Plane", which it says "provides a collaborative environment for humans and agents to scale intelligence with governance, oversight, and continuous calibration of enterprise knowledge".

A control plane is the part of a system that decides what is allowed. So it is worth asking the documentation what, specifically, it controls.

The term does not appear in the product documentation. "Semantic Control Plane" occurs zero times across Stardog's published documentation. It is a concept in a press release, and the press release says so: "the concept of". Governance and oversight are claimed in the announcement. The following six paragraphs are what the documentation describes.

There is no approve action. The published security model lists the complete, closed set of actions: read, write, create, delete, grant, revoke, execute, and "all", described as a "special action type that permits all previous actions over a resource". The words "approve" and "propose" do not appear on the page. Every change to a definition is a write, and every person who can make one can make it alone.

There is no approval to audit. Stardog's audit log is thorough about infrastructure, running from Authentication through WildcardPermissionRevoke. It contains no approval, review or submission event. The log will tell you that a definition changed. It cannot tell you that anyone agreed to it.

The human oversight that is documented is one person checking a machine. Voicebox offers "Human-in-the-Loop Oversight: Review mappings, resolve conflicts, and validate provenance", and the worked example has the same user review the AI's suggested ontology and then publish it. In Designer, the one place the product says "approve" is a mapping suggestion: "To approve a suggestion, click on the green check icon." That is real and useful, and it is a person checking an algorithm's guess. It is not a second person agreeing to a change.

The only approval state in the product belongs to another vendor. Knowledge Catalog can import lifecycle states of Draft, Published and Expired, along with a field documented as "the term's approval status", from Microsoft Purview. Those are Purview's approvals, read in as metadata. Stardog documents no approval status of its own.

The documented answer to change history is to use someone else's product. Designer exports a zip, and the documentation explains that "the contents can be used in version control systems to document data model and mapping changes". That is a reasonable engineering suggestion. It is also an admission: the history of your enterprise's agreed meaning lives in a Git repository maintained by whoever remembered to commit, outside the platform.

Version tracking tells you that something moved. It does not tell you what. Schema versioning computes "a 64-bit hash from the contents of the reasoning schema graphs", and "the hash value for a schema can be checked to determine if any of the associated named graphs have been modified". The documented retrieval returns the current hash. There is no retrieval of a historical version, no rollback, and no record of what changed or who changed it. Stardog does support binding a query to a named schema, which is pinning in the useful sense, with the condition that you maintain the older version yourself as separate named graphs. The platform will serve the version you preserved. It will not preserve one for you.

The capability used to be richer. Stardog 7.2.1 published a versioning API with Version, Tags and VersionFinder classes, including getNext and getPrevious navigation over "the version history". That package is absent from the Stardog 12.2.1 package list and class index. The most generous reading is that version history was not what customers were buying.

And read access to the model is not enforced where it counts. Stardog's documentation states that "named graph permissions do not affect the schema axioms used in reasoning, and every reasoning query will use the same schema axioms. This is true even though some users might not have been granted explicit read access to schema graphs." It immediately qualifies this: "However, non-schema axioms in those named graphs would not be visible to users without authorization." So the data stays protected and the model does not. Restrict a user from reading part of your ontology and the reasoner applies it on their behalf regardless. That is a defensible decision about inference consistency, and a strange property for a control plane.

The pattern is consistent. Stardog is genuinely strong at reaching data wherever it sits and answering questions across it, and it is positioned as though reaching data were the same as governing meaning. On the evidence of its own documentation, the control plane controls access. The agreement is left to you and a Git repository.

Which search terms map to which category?

The vocabulary buyers use does not line up with the categories, which is part of why shortlists get mixed.

How the common terms resolve

  • "Triplestore" is barely used on its own any more. It appears mostly in comparison form, as triple store versus graph database.
  • "Knowledge graph platform" is the live umbrella term, and it is used by products in at least three of the six categories. It tells you almost nothing about what a product does.
  • "Taxonomy management system" and "ontology management system" map cleanly to category four and almost nowhere else. These are the most precise terms in the space, and the fact that fewer vendors reach for them is informative.
  • "Semantic layer" is ambiguous and resolves to either category three or category six depending on who is speaking.

How do you tell which category you need?

Start from the problem you actually have. Four questions separate the categories faster than any feature grid.

Four questions that separate the categories quickly

  • Where does the data need to live? If it must stay in source systems, that points at category three. If you are consolidating into a graph, category one or two.
  • Is the hard part the query or the agreement? If teams disagree about what a term means, no amount of query performance fixes it, and the problem is category four.
  • Does a change need approving, and does the approval need to survive an audit? Ask to see propose and approve as separate permissions, the approval recorded with an author and a date, and the behaviour on deletion. Ask which licence tier each of those is in.
  • Who does the work day to day? Categories one, two and three are operated by engineers. Category four is operated by the subject matter experts who own the meaning, and the tooling has to suit people who will not be writing SPARQL.

What makes something an ontology management system?

Shipping a model editor does not make a product an ontology management system, any more than shipping a text box makes a product a content management system. The category is defined by what happens to a definition between someone proposing it and a downstream system consuming it.

Six requirements that define the category

  • Authoring by the people who own the meaning. Subject matter experts, in an interface built for them, without writing SPARQL or editing Turtle files.
  • Propose and approve as separate permissions. Not two people with write access and good manners. A documented permission to submit a change that is distinct from the permission to accept it.
  • An approval record that survives an audit. Who approved which change, when, with what comment, retrievable years later by someone who was not there.
  • Full change history on the definitions themselves, inside the product, queryable. Not an export you remembered to commit to a repository.
  • Named, citable versions a consuming system can bind to. A downstream application must be able to state which agreed version it is reading and keep reading it while the next one is drafted, without the team hand-maintaining a copy of the old one to point it at.
  • Deletion governed like any other change. Removing a concept from an approved model is the highest risk edit available and should be the hardest one to make unilaterally.

These are not advanced features. They are the minimum for treating meaning as a managed asset instead of configuration. A product that meets none of them is a database, whatever the homepage says, and a product that meets some of them behind a licence upgrade is a database with an upsell.

Where does TopQuadrant sit?

Category four, and the comparison that matters is against the other products in that category, and not against databases.

The distinction the category turns on is that the other five store graphs, serve queries or catalogue assets. Category four exists to hold the agreement: which definition is authoritative, who approved it, what it replaced, which parts of the business it binds, and which version a consuming system is reading. TQ Data Foundation was built for that question instead of being extended toward it, which is why governance is not a licence tier.

On portability, which this category argues about more than any other: Stardog's marketing promises delivery "without locking you into a proprietary" engine, so portability is not a claim only one vendor makes. Test it. Do not take it on trust from anyone, including us.

The right test is not whether the query language is open. It is whether the model, its history and its approvals can all leave together in a standard form. Those three are not equally portable. A set of RDF triples is the cheapest part of a semantic asset to move and the cheapest to rebuild. The expensive part is the decade of decisions behind it: which definition won, who signed off, what it replaced, why the 2023 version was retired. Where that record does not exist inside the platform, there is nothing to export and nothing to lock in, and a portability guarantee covering the triples is a guarantee about the easy half.

So what should you take away?

Three claims, each testable against the vendors' own published documentation.

The three claims this article makes

  • Most semantic platform shortlists mix categories, and the mixing is encouraged by vendors who benefit from appearing in a category they did not build for.
  • The categories differ by what they treat as the asset, and that is a more reliable guide than feature overlap or the words on the homepage.
  • If the problem is disagreement about meaning, and not access to data, it belongs in taxonomy and ontology management. Check that the governance is in the product itself, and not only in the enterprise tier, and check what happens when someone deletes something.

Where should you go next?

For the discipline itself, separate from the tooling, see taxonomy management and ontology management. For the difference between the underlying model types, see ontology vs taxonomy vs knowledge graph. For how the data and AI platform vendors segment, which this article does not cover, see the ontology market guide.

The categories matter because a shortlist built across them compares products that answer different questions. Decide the category first. Then compare inside it.

TopQuadrant sits in taxonomy and ontology management. We built for the agreement and its record, and that is a narrower claim than the one most of this market makes. Test it against our documentation and theirs.

Give your AI the context it's been missing

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