Snowflake for Gen AI Overview Study Guide — GES-C02

Snowflake for Gen AI Overview

This domain sets up everything else on the exam: where Cortex actually runs, why that matters for governance, the four building blocks you choose between for any given problem, how Snowflake's own AI agents differ from each other, and the managed-vs-self-hosted decision. Expect scenario questions that describe a business problem in plain language ("a business user wants to ask questions about sales data without writing SQL") and ask you to name the right Cortex capability — not syntax recall. Getting the capability map and the CoCo/CoWork/Cortex Agents distinctions solid is worth more than memorizing any single function signature.

Cortex Runs Inside the Account

Snowflake Cortex is Snowflake's suite of managed generative AI features: LLM functions, Search, Analyst, and Agents. The single fact the exam returns to repeatedly is that Cortex inference happens inside your Snowflake account's governance boundary:

• Prompts and data sent to a Cortex function are not shipped to a third-party LLM API the way a direct call to an external provider would be.
• Customer data is not used to train the underlying foundation models.
• Existing Snowflake access controls — roles, masking policies, row access policies — apply to the tables Cortex reads, exactly as they would for any other query.

This is why Cortex is the default answer whenever a question frames a scenario around regulated or sensitive data ("a healthcare company wants generative AI over patient notes without any data leaving Snowflake").

Serverless Compute and Credit Billing

Cortex LLM functions are serverless. You do not provision or size a virtual warehouse to run AI_COMPLETE or EMBED_TEXT_768 — Snowflake manages the underlying GPU/inference compute and meters usage as credit consumption, billed per token or per function call depending on the service. This is a different billing model from warehouse-second billing, and the exam likes to test the distinction: a virtual warehouse still needs to be running to issue the SQL statement and read/write table data, but the LLM inference step itself does not consume warehouse credits — it consumes separate Cortex serverless credits.

Cortex Search and Cortex Analyst also run as managed serverless services (Search additionally uses ongoing serverless compute to keep its index fresh).

The Cortex Capability Map

Four building blocks cover almost everything Cortex does, and this decision tree is the single most examinable idea in the whole domain:

Worked scenario: "Our support team wants to search five years of ticket transcripts for how a bug was resolved before." That is unstructured text retrieval → Cortex Search. "Our finance team wants to ask 'what was Q2 revenue by region' without writing SQL." That is a computed answer from structured tables → Cortex Analyst. "Our ops team wants one assistant that can look up a policy document and pull the matching account's order history and summarize both." That needs more than one tool orchestrated together → Cortex Agents.