AI in Snowflake: Snowpark and Cortex

By Kalle Siukola | October 8, 2026

Cortex provides a seamless way to use LLM models agains the data. User can select the model and the semantic model from the data is fetched automatically into the context. Non-technical users can use natural language to get information from the data.

With Snowpark, there is no need to use custom object relational models or construct dynamic sql queries. When building machine learning models or other data analysis, Snowpark provides similar API with the data than spark.

Each type of user has an entry point that matches the way they already work. All three paths call models that Snowflake hosts and read data under the caller's existing role, so no copy of the data leaves the platform.

Cortex: managed AI as a service

Cortex brings data available for users without deep technical knowledge. There are many capabilities to orchestrate and untilize Cortex, from simple we UI to code tools.

AI SQL functions. Use Snowflake AI Capabilities for a table data like any other transformations, native functions with a model selection.
Cortex Search. A managed hybrid search service over documents and text columns. It is easy to setup retrieval with indexing and vectorization without manual configurations. With semantic tables, consistent results can be achieved.
Cortex Analyst. Natural-language questions over structured data. Business users ask "what was revenue by region last quarter" and Cortex Analyst produces the SQL, guided by a semantic model that encodes the organisation's definitions and metrics.
Document AI. Extracts structured fields from invoices, contracts, and forms, and improves with a small number of labelled examples.
Snowflake Intelligence and Cortex Agents. A conversational interface and agent framework that combines Search and Analyst, so one question can draw on both documents and tables.
ML functions. Forecasting, anomaly detection, and classification exposed as SQL for teams that need predictions but not a custom model.

Snowpark: code where the data lives

Snowpark lets engineers and data scientists write Python, Java, or Scala that runs inside Snowflake's compute. A Snowpark DataFrame looks like pandas or Spark, but the work is pushed down to the warehouse, so a transformation over a billion rows does not require moving a billion rows anywhere. The same code can be packaged as a user-defined function or stored procedure and called from plain SQL by anyone in the organisation.

For heavier workloads, Snowpark Container Services runs user defined containers, including GPU workloads, inside the Snowflake account. This is how teams train custom models, host open-source models, or run existing Python services in one place inside Snowflake. Snowflake's Model Registry and Feature Store then handle versioning and deployment of the models that come out of that work.

Why this matters for a business

Data does not move. Role-based access, masking policies, and audit logging that already protect the warehouse apply to every AI call made against it.
No new infrastructure. There are no model servers, vector databases, or API gateways to run. Usage is metered like any other Snowflake compute.
The existing team can use it. Analysts stay in SQL, engineers use Python, and business users ask questions in plain language. Nobody needs to learn a new platform.
Choice of models. Cortex offers models from multiple vendors and sizes, so cost and quality can be balanced per use case.