AI & Tech

The W3C's Ontology Is for Inference; Palantir's Defines Actions

When every vendor says it has an ontology, what should a buyer ask? This piece reads the official documents through four questions: the unit you define, what the consuming side does with it, whether there is formal inference, and where actions live. It is a reading framework, not a test of what the products can do.

2026.09.29 · By dvdmaru · ~11 min read · 2,615 words

本文另有中文版:同叫 ontology:W3C 拿來推論,Palantir 把動作寫進定義

Say your team is building a system that looks up an order and then marks it “shipped.” Three vendors come in to pitch, and all three say they have an ontology. (This scenario is purely illustrative. It does not describe how any vendor’s product actually works.)

What would you want to ask? Probably this. What do I have to define first? Once it is defined, who uses it and for what? Will it infer conclusions I never wrote down? And where does the “mark as shipped” step live: in the definition, or somewhere else?

Those four questions are the method this piece uses to read official documentation. The same word means very different things in different documents, so the name alone does not tell you which one you are dealing with.

First, how messy the name is

Documents do not use the word tidily. Cube’s documentation calls its data model a knowledge graph. Microsoft’s Fabric IQ page calls its ontology a semantic modeling layer. Palantir’s architecture page says the Ontology is not a semantic layer, because a thin semantic layer cannot do the four-way integration of data, logic, action and security. The same page then uses the word semantic for the Ontology’s own parts, paired with kinetic.

So this is a comparison of what the documents say, not a test of what the products can do. The pages were fetched on September 29, 2026 and will change; where this piece says a page does not say something, it means the pages read do not say it, which is not the same as the product being unable to do it. This is part 4 of a series, after part 1, part 2 and part 3; it takes up the question part 3 ended on.

Four questions, plus permissions

Beyond the four questions in the opening, where permissions attach is a fifth, supporting question.

The fourth question does not ask whether there is an action, because Looker has an action parameter on fields and dbt has exports, so the answer would be yes for both. It asks at which level the action is written, which lets each product be put back in its own place.

Run through those questions, the documents center on three places. The grouping is the author’s organization, not a standard the W3C or the vendors share:

Basic unitWhat the consumer does with the definitionFormal inferenceWhere actions live
W3C (OWL 2)class, property, individual, axiomreasoners and other tools infer from the formal semanticsyes, centralpage does not say
Palantir documentationobject type, link type, action type, function and othersapplications, the OSDK, people and agents read and write objects and invoke actionsnone of the ten pages read use formal-inference vocabularythe action type is a type inside the definition
semantic layer family (dbt, Cube, LookML)metrics, dimensions, entities and joinsgenerate SQL; Cube validates a query and applies policies before it reaches the warehousepage does not sayLooker: a field parameter; dbt exports: a table in the same platform; Cube: page does not say

Looker and Fabric IQ do not fit the tidy cells of this table, and each gets its own section below.

The W3C’s ontology is a vocabulary for reasoners

This section covers OWL 2 as a standard, not every semantic web technology. The OWL 2 Overview defines ontologies as formal vocabularies of terms, typically covering a domain. The basic units are classes, properties and individuals, and the Primer calls the statements among them axioms.

What the consuming side does with the vocabulary is spelled out in the Primer: “Appropriate tools (so-called reasoners) can then be used to infer further information about that state of affairs.” On the third question the answer at this end is yes, and it is the core.

The Primer also draws lines so that readers do not mistake OWL for something else. “OWL 2 is not a database framework.” A database usually treats a missing fact as false, the closed-world assumption. OWL works under an open-world assumption, where a missing fact may simply not have been written down. OWL also does not assume that different names mean different individuals. It is not a syntax-checking schema language and cannot force a piece of data to appear. The Primer adds that a domain or range statement “is not a constraint on the knowledge, but allows a reasoner to infer further knowledge.”

As for actions, the W3C pages read describe vocabulary, graph data and validation, and say nothing about acting on external systems.

dbt, Cube and LookML define metrics and dimensions, then compile and restrict queries

dbt’s unit needs a version label. The block on the semantic models page is marked for v1.12 and later, and its example already uses simple metrics, so on that page the unit is the semantic model (with entities, dimensions and simple metrics) plus metrics. The MetricFlow page says that when a metric is requested it uses its own SQL engine to find the best path between tables from the YAML definitions. The definitions are used to plan SQL, and downstream tools query through GraphQL and JDBC.

Cube’s units are cubes and views, representing business entities like customers and orders, with measures, dimensions and joins. On the consumer question it is the most explicit: “Every query passes through the semantic layer runtime, where it’s validated against the data model and has access policies applied deterministically before reaching the warehouse.” The definition is a checkpoint on the query path. The knowledge graph mentioned earlier is, for Cube, this data model, and the model’s units are still cubes and views. That does not show whether Cube supports OWL-style inference.

LookML is the language Looker uses to build semantic data models, and Looker’s SQL generator turns it into SQL. Snowflake’s semantic view page mentions AI use, and the phrase “combining LLM reasoning with rule-based definitions” refers to LLM reasoning, which this piece reads as not being formal inference in the W3C sense.

Whether the consumer is an AI agent does not separate the documents, since Cube, Snowflake, Fabric IQ and Palantir all list agents among consumers. What separates them is what the definition is used for: compiling and restricting queries, inferring from axioms, or reading and writing objects and invoking actions.

Palantir’s documentation puts the action type in the definition

Palantir’s core page gives three units a sentence each: an object type defines an entity or event in an organization, a link type defines a relationship between two object types, and “An action type defines how an object type can be modified.” The action type is listed alongside the other two as a type inside the definition. The action-overview page says an action type “also includes the side effect behaviors that occur with action submission.” Action, rules and side effects sit in one definition, which part 2 of this series walked through, and side effects are optional.

For external systems, the webhooks page describes two types. With a writeback webhook, if the external request fails the Ontology applies no change; but the external request can succeed while the Ontology change fails, which the page describes as a degree of transactionality. A side-effect webhook runs after the object change, a failure is not shown to the end user, and the page says to use it for best-effort notifications or writing back to several external systems. So this documentation spells out its own limits on transactionality.

On the third question, start with a string check. The ten Palantir pages read do not contain OWL, RDF, SPARQL, SHACL, entailment or reasoner. “Reasoning” does appear, in the sense of decision logic or LLM reasoning. This is a limited string search of the pages fetched and cannot be written up as Palantir never supporting them.

Looker and dbt: actions exist, written differently

An objection comes up here. Doesn’t Looker have actions? Doesn’t dbt write tables? Both objections are backed by documentation, and the two are different kinds of thing.

Looker’s action parameter sits on a dimension or measure and lets users perform field-level tasks in other tools. The receiving server must accept a JSON POST, and a successful HTTP response counts as the action succeeding. So a Looker data action defines the request that is sent, and the receiving side carries it out. The Action Hub is a separate path, which the documentation itself keeps apart.

dbt exports write the results of a saved query to a table or view in the same data platform, as a leaf node of the DAG that other models cannot ref back. That is not an external action in the ordinary sense, and what dbt does toward external systems is not described in the pages read.

Together the two say one thing. In the semantic layer family, the definitions center on metrics, dimensions, entities and queries, and actions appear elsewhere in varying forms. Setting this beside Palantir compares how documents are structured and how much they disclose. It is not a verdict on capability.

Fabric IQ: the ontology label, with different documents emphasizing different things

The Fabric IQ ontology page was marked preview when read, and everything here refers to that state. It calls the ontology a business context and semantic modeling layer between enterprise data and its consumers, and says it can be built from scratch or generated from a semantic model. Its units are entity type, property, relationship, metric and rule, where a rule is a natural-language description of business logic. The page has no inference, reasoner or entailment wording.

On the relation to the W3C, the page says the ontology can import supported RDF, Turtle and OWL definitions and export supported RDF or Turtle formats. Importing OWL does not by itself show a built-in OWL reasoner. The graph is an optional execution layer, and the schema graph derived from an ontology does not load instance data by default.

For actions, name the document, because the subjects differ. The rules page’s subject is the rule: “Ontology doesn’t run the rule against data or execute actions.” This describes the new experience; the same September 29 snapshot already distinguished the new and old experience. The overview page’s subject is agents, which it says can trigger governed actions. In the September 29 snapshot, the June 2, 2026 Build announcement lists actions among the things ontologies define; the post does not detail the execution and authorization mechanisms for those actions.

Put the three documents together and you cannot fold them into “Fabric has no actions at all,” nor into “Fabric is the same type as Palantir.” On the unit and the consumer, the ontology resembles the semantic layer family. The page also has natural-language rules and RDF/OWL import, which the semantic layer pages do not mention.

Back at the buying table: what to ask, what to listen for

So when someone says “we have an ontology,” the name alone cannot answer your question. It is more useful to lay out the four questions:

You askYou listen for
What do I define first?classes, metrics and dimensions, object types, or something else
Once defined, who uses it, and for what?inference, compiled queries, or reading and writing objects and invoking actions
Is there formal inference?whether the documentation mentions reasoners and entailment, or LLM reasoning
Where do actions live?a type in the definition, a field parameter, a separate path, or not described
(Also) Where do permissions attach?on the definition, on the query path, or on the source data

In these technical reference pages, Palantir spells out in concrete terms how an action type modifies things and carries side effects, while Fabric’s various documents describe the definition, who executes, and the preview scope separately. Neither can be classified by whether it has actions. What separates Palantir from the semantic layer family is not the word semantic, which both use.

The most useful follow-up is the fourth question, pushed down to the page: where is that action defined, in which document, who runs it, and what happens when it fails. The name cannot answer those.

Next in this series: Palantir’s Row Permissions Stop at the Read; Markings Travel by Default.

FAQ

How do I tell which kind of ontology or semantic layer a vendor means?

Ask four questions of the documentation: what is the basic unit it tells you to define, what does the consuming side do with the definition, is there formal inference, and where do actions live. Add a fifth check on where permissions attach. This is the author’s way of reading official documents, not any vendor’s or standard’s classification, and not a test of product capability. The section “Four questions, plus permissions” shows how to use them.

Is Palantir’s Ontology the same thing as a W3C OWL 2 ontology?

The two sets of documents center on different things. The W3C’s OWL 2 documents describe an ontology as a vocabulary with formal semantics that reasoners can infer from. Palantir’s documentation has users define object types, link types, action types and functions. The ten Palantir pages read for this piece do not contain OWL, RDF or SHACL, but that is a limited string check of the pages fetched and cannot be written up as Palantir never supporting them.

Is Microsoft Fabric IQ’s ontology a semantic layer?

It depends on which question you ask. When read in September 2026 the page was marked preview, called the ontology a semantic modeling layer, said it can be generated from a semantic model, and listed metrics among its units. Those points resemble the semantic layer family. The page also has rules written in natural language and import of supported RDF, Turtle and OWL, and importing OWL does not mean a built-in OWL reasoner. Actions are spread over different documents with different subjects. This piece does not assign it a category.

Can a semantic layer only read data, never trigger an action?

The documents read here do not support that blanket claim. Looker has a field-level action parameter that is carried out by a receiving server, and dbt exports write query results to a table or view in the same data platform. They are different kinds of thing, and neither is the same as Palantir’s documentation treating the action type as a unit of definition.

Sources and scope

This piece compares what official documentation says, not what the products can do. Pages were fetched on September 29, 2026; editorially revised October 16, 2026, without incorporating later page changes.

  • W3C: OWL 2 Primer and Overview; the SHACL page (read in part).
  • Palantir (the ten pages of body text read): Why Ontology, Overview, Core concepts, Architecture (ontology-system), Object types, Action types, Webhooks, Side effects, Permissions and OSDK, plus the interoperability and export/import pages.
  • Semantic layer family: dbt (semantic models, MetricFlow, exports), Cube (intro, access policies), Looker (LookML, param-action, Action Hub, BigQuery write-back page), Snowflake semantic views overview.
  • Microsoft: the Fabric IQ ontology page (preview), the rules page, the overview page and the June 2, 2026 Build announcement. The Build announcement and rules page are quoted from snapshots fetched on September 29, 2026.
  • Pages not read, and nothing inferred from them: dbt’s GraphQL, JDBC, Python SDK and permissions subpages; Looker’s dedicated audit-log page; Databricks and Snowflake permissions subpages.