Data context and governance

    Semantics, ontology, context... buzzword bingo, or business value?

    A semantic layer defines how business numbers are calculated, an ontology defines the things a business deals with and how they relate, and a context layer maps where data lives and what it means. All three describe data without being data. People now use the terms interchangeably, which is how a team buys one and expects the other two.

    By Toby Mather · Published

    Key takeaways

    • A semantic layer defines metrics, an ontology defines business entities and their relationships, and a context layer maps systems, storage and meaning.
    • A knowledge graph stores entities and relationships as data, usually in a graph database, while a context graph is usually a graph of metadata built for agents to navigate.
    • The facts themselves, such as who the CFO is or what a customer paid, belong in the warehouse, where they can be governed and kept current.
    • These maps existed long before AI, and agents are the first readers that consult them on every question, which is where the business value comes from.
    • Treating the terms as one thing usually leaves agents with metric definitions and no idea which table or system to look in.
    1. 01Warehouse data

      The territory: orders, accounts, people

    2. 02Semantic layer

      How metrics are calculated

    3. 03Ontology

      Entities, relationships and rules

    4. 04Context layer

      Systems, storage and meaning

    5. 05Agents and reports

      Read the maps, then query the data

    Each map describes the warehouse data from a different angle. The facts stay in the warehouse, and agents read the maps to find them.

    Are these just buzzwords, or is there business value now?

    The ideas are old. Data teams have written metric definitions since the first OLAP cubes, ontologies come out of knowledge engineering and the semantic web, and every decent data team keeps some kind of map of its systems, even if it lives in one engineer's head.

    Agents have changed who reads the map. A dashboard user who knows the business can work around a missing definition, while an agent reads whatever description it is given on every question and guesses from column names when the map has a gap. A written, maintained map is now worth paying for in a way it rarely was when only analysts used it.

    The vocabulary has not kept up. Vendors use the same word for different products, and one product can get three names depending on who is selling it, so the first job on most projects is agreeing what each word means.

    Why are none of them data?

    Data is the record of what happened: an order, a payment, a contact's job title. Semantics, ontology and context describe that record. They say what an order is, how it relates to a customer, which table holds it and how revenue is computed from it. You could empty every table in the warehouse and the definitions would still make sense.

    The split matters because data and descriptions need different handling. Data changes every minute and needs access control down to the row, while descriptions change only when the business does and need an owner and a review step. Mixing the two, for example by writing customer facts into a context file, gives you facts with no access control and definitions nobody maintains.

    What is a semantic layer in this comparison?

    A semantic layer is the dictionary for business numbers: metric names, the SQL behind each one, the joins and the filters. It answers how the company calculates net revenue or churn, and very little else.

    Of the three, it has the most mature tooling. dbt, Snowflake, Databricks, Cube and Power BI all ship one, and the open Apache Ossie format describes metrics, dimensions and relationships so they can move between tools. Our semantic layer explainer covers it in full.

    What is an ontology, and how is it different?

    An ontology defines the kinds of things a business deals with and the rules between them. A customer has contracts, a contract has invoices, an invoice belongs to one legal entity, and a churned customer is one whose last contract ended without a renewal. Where a semantic layer defines numbers, an ontology defines the nouns, the relationships and the states.

    Palantir made the word familiar in enterprise software. Its Foundry Ontology maps datasets into object types, properties and links, and adds actions and functions so people can change data through the model. A mid-market company usually needs the same idea at a smaller scale, built over the warehouse they already run.

    What is a context layer?

    A context layer is the navigational map a person or an agent uses to find the right data. It records which systems exist, where their data lands in the warehouse, what each table and column means, which tables are trusted, how metrics are defined, and the caveats a new analyst would otherwise learn the hard way. In practice it contains the semantic layer and the ontology, plus everything else needed to get from a question to the right table.

    It is still a map. The sentence "John is the CFO and the main buyer on the Acme deal" is a fact about the territory, and it belongs in a table in the warehouse, where the CRM sync keeps it current. The context layer tells an agent which table to look in. Our post on context covers this in more depth.

    Where does a knowledge graph, or a context graph, fit?

    A knowledge graph stores entities and the relationships between them as data, usually in a graph database such as Neo4j. It is good at questions that follow links, such as which suppliers share a director or which documents mention a product, and search and recommendation systems use them heavily.

    Context graph is a newer term, and vendors use it in different ways. DataHub describes its platform as a context graph that brings lineage, quality and real-time context together for agents. Atlan describes a context layer of metadata, lineage, glossary and governance that it serves to agents. Both are describing a graph of metadata about your data.

    The useful test is to ask what the nodes are. If the nodes are tables, metrics and systems, you have a map. If the nodes are customers and deals, you have business data stored as a graph, and it needs the same access control as any other customer data.

    TermWhat it definesTypical homeWhat stays in the warehouse
    Semantic layerMetric names, SQL, joins and filtersdbt, a warehouse, a BI tool or Rig MapThe numbers themselves
    OntologyEntities, relationships, states and rulesPalantir Foundry, a modelling tool or Rig MapThe customer and contract records
    Context layerSystems, storage, table meanings, trust and caveatsA catalog, a context platform or Rig MapFacts about deals and people
    Knowledge graphEntities and relationships stored as dataA graph databaseThe source records it was built from

    What goes wrong if you treat them as one thing?

    These are the failures we see when the terms get blurred together.

    • A team buys a semantic layer for its agents, and the agents still pick the wrong table for anything that is not a predefined metric
    • A team draws an ontology in a workshop with nothing connecting it to the tables, so it is out of date by the next schema change
    • Customer facts end up in context files or prompts, outside the warehouse's access control, and an agent repeats a job title that changed months ago
    • A catalog full of metadata gets mistaken for metrics that can be queried under policy

    Why do AI agents need all of them?

    Take an agent answering how enterprise churn moved last quarter. It needs the churn definition from the semantic layer, the meaning of an enterprise customer from the ontology, and the context layer to know that segment lives on the account table synced from Salesforce, where the billing system has a different customer ID. Missing any one of those, it guesses.

    It also needs permission to read those rows and nothing else. None of the three maps answers that question, which is why governed access sits alongside them as a separate layer.

    How Rig does it

    Rig Map holds the ontology, the semantic layer and the context layer in one place, over the warehouse you already run, and keeps them current as schemas change. Context Studio is where a data team curates table descriptions, usage rules, certified metrics and gold-star tables, grouped into business areas. The Systems Map shows how data flows between systems, from a Salesforce field to the report or workflow that uses it.

    The facts stay in the warehouse. Ingest, or the pipelines you already run, keeps the tables current, and Rig Map points at them.

    Gateway controls who reaches that data through the map. People and agents get roles and row-level access, each agent has its own token and daily call limit, and every call is logged under the agent's name.

    Tools and approaches

    Palantir Foundry Ontology
    Palantir describes the Ontology as an operational layer over datasets, virtual tables and models. It maps data into object types, properties and links, with actions and functions for changing data and running business logic.
    Atlan
    Atlan describes its context layer as the governed layer that connects AI agents to trusted data. It exposes metadata, lineage, glossary and governance to agents, with MCP as the delivery mechanism, without moving the data.
    DataHub
    An open-source metadata platform that describes itself as a context platform for AI agents. DataHub Core is the open-source edition and DataHub Cloud the commercial one, and it is MCP-native.
    Neo4j
    A native graph database for storing and querying connected data. Teams use it to build knowledge graphs that link structured and unstructured data.
    dbt Semantic Layer
    Metrics defined in the dbt project and served through MetricFlow to downstream tools, APIs and the dbt MCP server.
    Snowflake semantic views
    Schema-level objects in Snowflake holding logical tables, relationships, facts, dimensions and metrics, used by Cortex Analyst and Cortex Agents.

    Common questions

    Can I ignore these terms and just use a catalog?

    A catalog records what data exists, who owns it and where it came from, which covers a large part of a context layer. Most catalogs do not hold metrics you can query under policy, and they do not control what an agent is allowed to read. If you use one, connect it to a semantic layer and a governed access path.

    Is a context graph just a knowledge graph with better marketing?

    It depends on what the nodes are. A context graph whose nodes are tables, metrics and systems is a metadata map. One whose nodes are customers and deals is a knowledge graph of business data, and it needs the same access control as the warehouse.

    Do I need all of them?

    You need the parts your agents and reports actually use. We usually start with certified metrics for the numbers people argue about, add table descriptions and relationships for one business area, and grow the map from there.

    Where does Palantir's Ontology sit in this picture?

    Palantir's Foundry Ontology maps datasets into object types, properties and links, and adds actions that change data. It is an ontology with an operational layer on top. Rig Map applies the same idea for mid-market companies on the warehouse they already run.

    Which one should agents use?

    All three, through one access point. The context layer tells the agent where to look, the ontology tells it what the things are, and the semantic layer gives it the agreed calculation. A gateway decides what it may read.

    Sources

    Vendor documentation checked on 20 November 2026. Confirm current features and terms with the vendor.

    About the author

    Toby Mather, Co-founder and CEO, Rig. Toby co-founded Rig after seeing how hard it was for business teams and AI to work with complex internal data. He was previously co-founder and CEO of Lingumi and CPO at Novakid.

    Talk to Rig

    See your stack drawn as one diagram

    An architect drafts a one-page diagram of your warehouse, sources and access path, then offers a 30-minute walkthrough.