Key takeaways
- A semantic view holds logical tables, relationships, facts, dimensions and metrics as one schema-level object, with synonyms and comments for business language.
- You query it with the SEMANTIC_VIEW clause in SQL, and Cortex Analyst and Cortex Agents read its definition to generate SQL against the physical tables.
- Creating one needs CREATE SEMANTIC VIEW on the schema and SELECT on the underlying tables, and masking and row access policies still apply at query time.
- Semantic views cover defined metrics on Snowflake data well, while systems outside Snowflake, table meanings beyond the model and per-agent identity need other layers.
- Our short guide to creating a Snowflake semantic view covers the setup steps, and this post covers where semantic views stop.
- 01Snowflake tables
Data synced from your systems
- 02Semantic views
Metrics, dimensions, relationships
- 03Context layer
Systems, table meanings, usage rules
- 04Gateway
A role, token and log per agent
- 05AI agents
Cortex, Claude, Cursor
What is a Snowflake semantic view?
Snowflake describes a semantic view as a schema-level object that stores business meaning alongside the physical data. It contains logical tables, which usually correspond to business entities such as customers or orders, and relationships that join those tables on shared keys. Facts are row-level attributes used to build other concepts, dimensions are the categories you group and filter by, and metrics aggregate data into measures of business performance.
Snowflake treats semantic views as metadata. They sit in a schema next to your tables and views, and you can share them in private listings, organisational listings and public Snowflake Marketplace listings.
What do you put in one?
A first semantic view usually covers one business area. For revenue that means logical tables for customers, contracts and invoices, the relationships between them, dimensions such as region and segment, and metrics such as net revenue and average contract value. Logical tables, facts, dimensions and metrics all accept synonyms and comments, which give Cortex Analyst the words people actually type.
You can write the CREATE SEMANTIC VIEW statement by hand, build the view in Semantic Studio, or use Semantic View Autopilot in Snowsight, which accepts YAML, Tableau or Power BI files as a starting point. Our setup guide walks through creating one step by step.
How do access and sharing work?
Semantic views follow ordinary Snowflake RBAC. Creating one needs the CREATE SEMANTIC VIEW privilege on the schema and SELECT on every table and view it uses. Snowflake enforces masking and row access policies in the query engine, so they apply to people, BI tools and AI agents alike.
Horizon Catalog is the wider catalog these objects sit in. Snowflake describes Horizon Context as the governed context layer inside Horizon Catalog, made up of semantic views, lineage, auto-generated descriptions and object tagging. If all of your data is in Snowflake, that covers a good share of a context layer before you add anything else.
How do Cortex Analyst and other agents use them?
Cortex Analyst takes a semantic view in its REST API request and generates SQL from a natural-language question. Cortex Agents can add a semantic view as a tool, and Snowflake notes that the agent currently reads the definition in the semantic view and generates SQL against the physical tables. The Snowflake-managed MCP server can expose Cortex Analyst over semantic views to outside MCP clients such as Claude.
Snowflake's own MCP documentation warns that exposing the SQL execution tool alongside an agent lets clients bypass the agent's semantic views. That warning applies to every setup: a semantic view only helps when the agent is made to go through it.
Where do semantic views stop?
Semantic views are good at defined metrics on Snowflake data. The rest of the job sits in a context layer and an access layer.
- Questions that no metric covers, such as tracing one order or checking why a payment failed, need table meanings and joins beyond the model
- A CRM field that never synced, a workflow inside a billing tool or a spreadsheet the finance team owns does not appear in a semantic view
- RBAC acts on the role a session uses, so if every agent connects through one service role, the query history and the row policies see one identity
- A semantic view records the definition but not who approved it or what changed last month, so it needs a review process around it
- A metric that exists only as a Snowflake object is read by Snowflake and its clients, which matters after an acquisition brings a second warehouse or a BI tool outside Snowflake
How does this relate to a context layer?
A semantic view is the metric part of the map, for data that lives in Snowflake. A context layer adds the systems that feed Snowflake, what each source table means, which tables are trusted, how customer IDs match across systems, and the usage rules an analyst would tell a new hire. Horizon Context covers part of this for Snowflake objects, and teams whose estate includes CRMs, ERPs and other warehouses usually need the map to reach further.
Catalog vendors take a different route. Atlan documents a context layer of metadata, lineage, glossary and governance that it serves to agents, often over MCP. That approach sits on the catalog side, and some teams run it next to semantic views.
Why do AI agents need more than a semantic view?
An agent with a good semantic view answers defined metric questions well. Ask it why enterprise churn rose in Germany and it needs much more: the map to find the account table, the support tickets and the contract changes, and the right to read those rows as itself. Each piece covers a different failure, and an agent missing any of them goes back to guessing.
Should metrics live in Snowflake, dbt or somewhere else?
If Snowflake is your only warehouse and your agents are mostly Cortex, putting metrics in semantic views keeps them next to the data and its policies. If you already run the dbt Semantic Layer, or expect to run more than one engine, choose one source of truth and generate the others from it, so two copies of revenue never drift.
Apache Ossie, the open semantic format that Snowflake helped found under its earlier name of Open Semantic Interchange, is designed to make that kind of exchange easier. Our dbt comparison post goes through the trade-offs in more detail.
How Rig does it
Snowflake is one of the warehouses Rig runs on. Rig Map holds certified metrics, table descriptions, usage rules and the Systems Map over your Snowflake data, in an OSI-compatible format that exports as a pull request for review, and it maps the systems that feed Snowflake as well as the tables inside it.
Gateway gives each agent its own role, token and daily call limit over that data, applies row scoping, and logs every call under the agent's name. Birdie runs this way on its own Snowflake, with its metric definitions and dbt models moved in as certified metrics.
Tools and approaches
- Snowflake semantic views
- Schema-level objects for logical tables, relationships, facts, dimensions and metrics. Created with SQL, Semantic Studio or Semantic View Autopilot, and used by Cortex Analyst and Cortex Agents.
- Snowflake Horizon Catalog
- Snowflake's catalog for data inside and outside Snowflake, covering discovery, lineage, data quality, masking and row access policies. Horizon Context is its governed context layer.
- Snowflake-managed MCP server
- Exposes Cortex Analyst over semantic views, Cortex Search, Cortex Agents and other tools to MCP clients, with a separate grant needed for each tool.
- Atlan
- A catalog and governance platform whose context layer serves metadata, lineage, glossary and governance to AI agents, with MCP as the delivery mechanism.
- dbt Semantic Layer
- Metrics defined in the dbt project and served through MetricFlow to several tools.
- Cube
- A separate semantic layer with metrics, dimensions, joins and access rules, served to AI clients over MCP.
Common questions
Are semantic views a context layer on their own?
They are the metric part of one. Snowflake groups semantic views with lineage, descriptions and tags under Horizon Context. A full context layer also covers systems outside Snowflake and the rules for using each table.
How do semantic views differ from ordinary views?
An ordinary view fixes its query when you create it. A semantic view defines metrics, dimensions and relationships, and you pick which metrics and dimensions to combine at query time with the SEMANTIC_VIEW clause.
Can I use semantic views alongside dbt?
Yes, and the main decision is which one holds the source of truth for each metric. If the dbt Semantic Layer already defines revenue, generate or check the semantic view from it instead of maintaining two versions by hand.
Do semantic views replace row access policies?
Semantic views define meaning, and row access and masking policies decide what each role can see. Snowflake enforces those policies in the query engine for whichever role runs the query, so you still need them.
How do I create a Snowflake semantic view?
Use CREATE SEMANTIC VIEW in SQL, Semantic Studio, or Semantic View Autopilot in Snowsight. Our setup guide for Snowflake semantic views walks through the steps.
Sources
Vendor documentation checked on 20 November 2026. Confirm current features and terms with the vendor.