CertSafari
    Snowflake SnowPro Specialty: Gen AI (GES-C02)· Lessons

    Domain 1 · Lesson 1/15

    Snowflake Gen AI principles and the Cortex feature map

    Define Snowflake’s Gen AI principles and features.

    19 min read
    9% of exam
    10 sources
    Published 21 Oct 2026
    Docs as of 18 Oct 2026

    What you will be able to do

    • Explain how Snowflake keeps Gen AI processing inside its own governance perimeter, and say what that means for model training and RBAC
    • Pick the right Cortex feature for a use case: AI Functions, Cortex Search, Cortex Analyst or Cortex Agents
    • Describe Snowflake Intelligence and how it relates to Cortex Agents
    • Explain what Cortex Fine-tuning produces and what training data it needs
    • Decide when to bring your own model through the Model Registry and Snowpark Container Services instead of using a Cortex-managed model
    • Match each way of reaching Cortex (AI & ML Studio, SQL, Python, REST) and each coding assistant (Cortex Code, Snowflake Copilot) to the user it serves

    Key concept

    AI runs where the data lives — Snowflake brings the models to your governed data, not your data to the models. Cortex LLMs run inside Snowflake's service perimeter, customer data is never used to train models shared with other customers, and your existing roles and privileges still decide what any AI feature can read.

    1.The principle: models come to governed data

    Snowflake Cortex is Snowflake's suite of AI features built on large language models. Its design rests on one rule: the model runs where the data already lives. Cortex AI Functions give access to LLMs from OpenAI, Anthropic, Meta, Mistral AI, DeepSeek and xAI. Snowflake does not ship a copy of your rows to those vendors' infrastructure. It deploys the models inside the Snowflake service perimeter. When AI_COMPLETE summarises a support ticket, the ticket text never leaves Snowflake for an outside provider.

    The second half of the principle concerns training. Snowflake does not use Customer Data to train or fine-tune any model offered across its customer base. Your prompts don't improve a shared foundation model, and nothing you send can show up in another account's answers. Cortex Analyst shows the same idea from another angle. For inference it sends the model only semantic-model metadata, such as table names, column names and descriptions. The SQL the model writes then runs in your own virtual warehouse.

    The third half is governance. AI features don't get a side door around RBAC. SQL generated by Cortex Analyst still obeys every access control you have set up, and a Cortex Agent reads data only through Snowflake privileges and the execution context of each tool. If a role can't SELECT from a table, an AI feature acting for that role can't read it either. That is why an organisation can switch Cortex on without redesigning its security model.

    Checkpoint 1 of 9· Check yourself

    A company turns on Cortex Analyst. A business user's role has no SELECT privilege on the HR salary table. The user asks Cortex Analyst a question whose answer depends on salary data. What happens?

    Sources1234

    2.Choosing the Cortex feature: functions, Search, Analyst or Agents

    Four Cortex features cover most Gen AI work in Snowflake, and each has a different job. Picking one starts with a single question: what is the input, and what should come back?

    Which Cortex feature fits which job
    FeatureWhat it doesPick it when
    AI Functions (AI_COMPLETE, AI_CLASSIFY, AI_FILTER, AI_EXTRACT, ...)Run an LLM task on each row or file inside a SQL queryYou want to transform, classify, extract from or summarise data in bulk
    Cortex SearchManaged hybrid (vector and keyword) search over textYou need retrieval: a RAG chatbot, enterprise search, or document retrieval for an agent
    Cortex AnalystTurns natural-language questions into SQL over semantic viewsBusiness users ask questions about structured data such as sales or revenue
    Cortex AgentsPlans, calls Analyst, Search and other tools, then reflects on the resultsOne request needs several steps or both structured and unstructured data
    Snowflake Model Registry with Snowpark Container ServicesHosts a model you built yourselfNo Cortex-managed model fits and you have your own

    AI Functions are SQL functions that transform data. AI_COMPLETE handles most generative tasks. Task-specific functions handle the routine jobs: AI_SENTIMENT, AI_TRANSLATE, AI_REDACT for PII, AI_EMBED, AI_TRANSCRIBE for audio and video, and AI_PARSE_DOCUMENT for staged documents. AI_AGG and AI_SUMMARIZE_AGG work across many rows and aren't limited by a context window. All of these process inputs. None of them is a search index or a question-answering service over your schema.

    Cortex Search is the retrieval engine. It gives you a hybrid search engine over your text data without you managing embeddings, infrastructure, quality tuning or index refreshes. You define a service over a table, a view or (in preview) staged files. It refreshes itself according to a TARGET_LAG. Its two main use cases are RAG and enterprise search, and it also serves as the retrieval layer for Cortex Agents. In a RAG chatbot over thousands of PDF manuals, Search finds the relevant passages and an LLM function writes the answer from them.

    Cortex Analyst handles text-to-SQL. Business users ask a question such as total sales by region last quarter and get a direct answer without writing SQL. A plain schema doesn't tell a model how a metric like revenue is calculated, so Analyst relies on Semantic Views. These are schema-level objects that define logical tables, dimensions, facts, metrics and join relationships. Analyst is offered as a REST API, so it can sit behind a Streamlit app, Slack or Teams. The two features are easy to mix up. Analyst answers questions about structured data by generating SQL. Search retrieves passages from unstructured text. Neither one does the other's job.

    Behind every feature sits a choice of model. Agents accept auto, which picks the highest-quality model available to your account. Models also have lifecycle dates: once a model reaches its legacy date, only accounts that already used it can keep calling it. SHOW CORTEX BASE MODELS lists those dates so you can move to a newer model in time.

    Checkpoint 2 of 9· Match them up

    A data team has four requests. Match each Cortex feature to the request it is built for.

    Tap a term, then the definition that fits it.

    Sources1523

    3.Cortex Agents and Snowflake Intelligence

    Cortex Agents orchestrate the other features. An agent is a schema-level object that bundles a model, tools and natural-language instructions. Each request goes through a loop. The agent plans, which can mean clarifying a vague question or splitting it into subtasks. It then calls tools, such as Cortex Analyst for structured data and Cortex Search for unstructured data. Finally it reflects on the results and either calls another tool, asks a follow-up question or answers. Threads keep conversation context between turns, so your client application doesn't have to store it.

    Tools are the agent's only access to your data. An agent with no tools can still chat, but it can only use the model's general knowledge. To let an agent use an existing search service next to a semantic view, you add both as tools and configure the resources each one needs: the semantic view, the search service and a warehouse. Other tool types include a sandboxed Python code-execution tool, Data to Chart, custom tools built from stored procedures or UDFs, MCP connectors and web search.

    Snowflake Intelligence is the ready-made front end for people who will never call an API. It is an agentic application with a conversational interface that lets business users ask questions in natural language over both structured and unstructured enterprise data. It runs on data agents, keeps answers traceable, respects Snowflake's governance policies, and reduces how often users have to wait on the data team. Current Snowflake documentation lists this application as Snowflake CoWork. The split to remember: developers build Cortex Agents and call them through the agent:run REST API, while business users talk to those agents in Snowflake Intelligence or in Cortex Code.

    Checkpoint 3 of 9· Put it in order

    Put in order the steps a Cortex Agent follows when it receives the request "What are the differences between contract terms for Acme Supplies and Acme Stationery?"

    1. 1.Use tools: call Cortex Search to retrieve the relevant contract clauses for each supplier
    2. 2.Reflect: evaluate the tool results and decide whether to call another tool, ask a clarifying question or answer
    3. 3.Respond: generate the final answer that compares the two sets of contract terms
    4. 4.Plan: parse the request, split it into subtasks and choose a tool for each part

    Sources36

    4.Cortex Fine-tuning: adapting a base model to your task

    Fine-tuning fits the gap between prompt engineering and training a model from scratch. It is worth trying when prompting or RAG doesn't give good enough results or latency, and building a large model yourself would cost too much. Cortex Fine-tuning uses parameter-efficient fine-tuning (PEFT). It doesn't retrain the whole base model. It trains a customised adaptor that sits on top of a supported pre-trained model, such as llama3.1-8b, and the training runs entirely inside Snowflake.

    The training data comes from a table or view queried with SQL. The query result must contain two columns named prompt and completion. Your source columns don't need those names, because a column alias in the query is enough. The function ignores any other columns. Start with a few hundred examples: many more can greatly increase tuning time while improving quality only a little.

    Starting a fine-tuning job, using aliases to supply the required prompt and completion columnssql
    SELECT SNOWFLAKE.CORTEX.FINETUNE(
      'CREATE',
      'my_tuned_model',
      'llama3.1-8b',
      'SELECT a AS prompt, d AS completion FROM train',
      'SELECT a AS prompt, d AS completion FROM validation'
    );

    FINETUNE takes four actions: CREATE, SHOW, DESCRIBE and CANCEL. Jobs run for a long time and aren't attached to a worksheet session, so closing the tab doesn't stop one. Use DESCRIBE with the job ID to check progress. The result is a model object in a schema you choose. That is why the creating role needs CREATE MODEL on that schema, and other roles need usage on the model before they can call it. You call the tuned model by name with AI_COMPLETE, like any other model. You pay for training tokens (input tokens multiplied by epochs), then for inference tokens, plus storage for the adaptor.

    Checkpoint 4 of 9· Fill the gap

    Your training table keeps the questions in column a and the expected answers in column d. You don't want to rename any table columns. Which alias goes in the blank so that FINETUNE accepts this job?

    SELECT SNOWFLAKE.CORTEX.FINETUNE(
      'CREATE',
      'my_tuned_model',
      'llama3.1-8b',
      'SELECT a AS  ? , d AS completion FROM train',
      'SELECT a AS prompt, d AS completion FROM validation'
    );

    Sources7

    5.Bringing your own model: Model Registry and Snowpark Container Services

    Some models fall outside Cortex: a model you trained yourself, or one with an architecture Cortex doesn't offer. Snowflake ML covers that case, letting you develop and run custom models while the data stays in Snowflake. You log the model in the Snowflake Model Registry and call it from there, as covered in Snowflake Model Registry: logging and calling models. If an application needs low latency, the registry can deploy the model as a managed service with its own HTTP endpoint. Snowflake hosts that service in Snowpark Container Services (SPCS), where it can run on distributed GPU clusters without you managing Docker images or Kubernetes.

    The decision rule moves from least effort to most:

    1. If a Cortex-managed model can do the task through a prompt, use AI Functions. Snowflake hosts, scales and updates the model for you.
    2. If a supported base model is close but needs your domain's examples to perform well, use Cortex Fine-tuning. You still don't run any infrastructure.
    3. If the model is your own, or needs its own endpoint and GPUs, register it in the Model Registry and serve it on Snowpark Container Services.

    The registry is the right answer when you bring the model. It is not a catalogue of Cortex's own LLMs, and it is not a retrieval or text-to-SQL feature.

    Checkpoint 5 of 9· Check yourself

    Prompting llama3.1-8b through AI_COMPLETE gives answers that are close to what an insurance team needs. Adding retrieved passages with RAG has not improved them enough. The team has a few hundred labelled examples from its own domain and does not want to run any model infrastructure. Which option fits best?

    Sources48

    6.Ways to reach Cortex: AI & ML Studio, SQL, Python and REST

    The same Cortex features can be reached from several surfaces, and each one suits a different user:

    • AI & ML Studio in Snowsight is the no-code path. Under AI & ML you can create a Cortex Search service through a wizard, or start a fine-tuning job from AI Studio by picking a base model, a training table and the prompt and completion columns.
    • SQL is the main surface. Cortex features are SQL functions, so analysts and data engineers can call them inside SELECT statements, pipelines and WHERE clauses.
    • Python exposes the same features to data scientists and application code.
    • REST APIs serve applications that need responses quickly: the Complete API for inference, the Embed API for embeddings, and the Agents API, along with the Cortex Analyst and Cortex Search endpoints.

    Choose between SQL and REST based on the workload's shape. AI Functions are optimised for throughput: they suit a nightly job that classifies millions of rows in a table. A chat window where a user is waiting for each reply depends on latency, and that is what the REST API is for. The Studio wizard has limits of its own. If a search service needs several data sources or a transformation, define it in SQL.

    Checkpoint 6 of 9· Check yourself

    A support portal must show an LLM-written reply while the customer waits on the page. A separate nightly job labels the day's tickets in a table. Which pairing of surfaces fits?

    Sources17

    7.Coding assistants: Cortex Code and Snowflake Copilot

    Cortex Code (shown in the product as CoCo) is Snowflake's AI agent for building things: SQL, Python notebooks, pipelines, ML work and agents. It is aware of your RBAC and schemas. It comes in three forms. Cortex Code in Snowsight is built into Workspaces and the admin pages. It uses the SQL file or notebook you have open as context, shows suggested changes in a diff view for you to accept, and answers admin questions about credits, query performance and permissions. A desktop IDE for macOS and Windows offers the same agent with local file access. The Cortex Code CLI is for developers who work in a local shell.

    The CLI can do things the browser can't. It reads and writes your local repositories, which suits dbt projects and Streamlit apps, and it runs bash commands, git operations and SQL against your warehouse. You can customise it with AGENTS.md files, skills and MCP servers. Several safeguards limit what it does: Snowflake RBAC, Restricted Session Scope, OS-level sandboxing and a three-tier approval system. As a rule of thumb, an analyst in a worksheet uses Snowsight, and an engineer managing a repository uses the CLI.

    Snowflake Copilot is the older assistant. It is an LLM-powered helper for data analysis that runs on a Snowflake-fine-tuned model inside Cortex, so it follows the same rule of keeping data in Snowflake. Copilot works in worksheets in two ways: through a chat interface and inline. It runs natively in a limited set of AWS and Azure regions. To use it in other regions, set CORTEX_ENABLED_CROSS_REGION. Snowflake is replacing Copilot with Cortex Code, so new work should start with Cortex Code.

    Checkpoint 7 of 9· Exam question

    A prospective customer asks how Snowflake protects their data when using Cortex AI functions such as AI_COMPLETE. Which statement accurately describes Snowflake's core Gen AI security principle?

    Checkpoint 8 of 9· Exam question

    A retail analytics team wants business users to type plain-English questions like 'What were total sales by region last quarter?' and receive accurate SQL-generated answers against a governed semantic view. Which Snowflake Gen AI feature should they implement?

    Checkpoint 9 of 9· Exam question

    A support organization has thousands of unstructured PDF manuals and wants a fully managed retrieval layer that combines semantic vector matching with keyword relevance for a RAG-based chatbot. Which feature best fits this requirement?

    Sources910

    Sources

    (10)

    Every claim above is drawn from one of these pages, quoted as it was written on the date shown.

    1. 1.
      “All the LLMs that Snowflake provides access to via our Snowflake AI Features are deployed within the Snowflake Service perimeter.”
      ↩︎ The principle: models come to governed data
      “AI_COMPLETE: Generates a completion for a given text string or image using a selected LLM.”
      ↩︎ Choosing the Cortex feature: functions, Search, Analyst or Agents
      “Snowflake Cortex features are provided as SQL functions and are also available in Python.”
      ↩︎ Ways to reach Cortex: AI & ML Studio, SQL, Python and REST
      “For more interactive use cases where latency is important, use the REST API.”
      ↩︎ Ways to reach Cortex: AI & ML Studio, SQL, Python and REST
      “Cortex AI Functions are optimized for throughput.”
      ↩︎ Ways to reach Cortex: AI & ML Studio, SQL, Python and REST
      “All the LLMs that Snowflake provides access to via our Snowflake AI Features are deployed within the Snowflake Service perimeter.”
      ↩︎ Key concept
    2. 2.
      “We do not use your Customer Data to train or fine-tune any Model to be made available for use across our customer base.”
      ↩︎ The principle: models come to governed data
      “This SQL query is then executed in your Snowflake virtual warehouse to generate the final output.”
      ↩︎ The principle: models come to governed data
      “ensuring that SQL queries generated and executed adhere to all established access controls”
      ↩︎ The principle: models come to governed data
      “business users can ask questions in natural language and receive direct answers without writing SQL”
      ↩︎ Choosing the Cortex feature: functions, Search, Analyst or Agents
      “We do not use your Customer Data to train or fine-tune any Model to be made available for use across our customer base.”
      ↩︎ Exam trap 1
      “Cortex Analyst uses Semantic Views to understand your data and generate accurate SQL queries.”
      ↩︎ Exam trap 2
      “Cortex Analyst fully integrates with Snowflake’s role-based access control (RBAC) policies, ensuring that SQL queries generated and executed adhere to all established access controls.”
      ↩︎ Checkpoint
    3. 3.
      “Data access is governed by Snowflake privileges and the execution context of each configured tool.”
      ↩︎ The principle: models come to governed data
      “Run SHOW CORTEX BASE MODELS to see legacy and end-of-life dates”
      ↩︎ Choosing the Cortex feature: functions, Search, Analyst or Agents
      “When creating an agent, we recommend selecting auto for the model.”
      ↩︎ Choosing the Cortex feature: functions, Search, Analyst or Agents
      “They generate SQL over structured data using Cortex Analyst semantic views and use Cortex Search to retrieve insights from unstructured sources”
      ↩︎ Cortex Agents and Snowflake Intelligence
      “Add tools: Configure the tools the agent can use and the resources each tool needs, such as semantic views, search services, and warehouses.”
      ↩︎ Cortex Agents and Snowflake Intelligence
      “Tools that reach external services, such as web search and MCP connectors, send requests over the public internet.”
      ↩︎ Exam trap 5
      “They generate SQL over structured data using Cortex Analyst semantic views and use Cortex Search to retrieve insights from unstructured sources”
      ↩︎ Checkpoint
      “Tools that reach external services, such as web search and MCP connectors, send requests over the public internet.”
      ↩︎ Prediction
      “The LLM-driven plan, use tools, reflect loop the agent runs to answer a request.”
      ↩︎ Checkpoint
    4. 4.
      “Snowflake Cortex is a suite of AI features that use large language models (LLMs) to understand unstructured data, answer freeform questions”
      ↩︎ The principle: models come to governed data
      “Snowflake ML lets you develop and operationalize custom models to solve your unique data challenges, while keeping your data inside Snowflake.”
      ↩︎ Bringing your own model: Model Registry and Snowpark Container Services
    5. 5.
      “The two primary use cases for Cortex Search are retrieval augmented generation (RAG) and enterprise search.”
      ↩︎ Choosing the Cortex feature: functions, Search, Analyst or Agents
      “a hybrid (vector and keyword) search engine on your text data in minutes”
      ↩︎ Choosing the Cortex feature: functions, Search, Analyst or Agents
      “without having to worry about embedding, infrastructure maintenance, search quality parameter tuning, or ongoing index refreshes”
      ↩︎ Exam trap 3
    6. 6.
      “Snowflake CoWork is a ready-to-use agentic application with an intuitive, conversational interface that helps business users discover and act on deep insights.”
      ↩︎ Cortex Agents and Snowflake Intelligence
      “Unified data access: Analyzes both structured and unstructured data from enterprise sources.”
      ↩︎ Cortex Agents and Snowflake Intelligence
    7. 7.
      “Cortex Fine-tuning allows users to leverage parameter-efficient fine-tuning (PEFT) to create customized adaptors for use with pre-trained models on more specialized tasks.”
      ↩︎ Cortex Fine-tuning: adapting a base model to your task
      “the query result must contain columns named prompt and completion”
      ↩︎ Cortex Fine-tuning: adapting a base model to your task
      “Fine-tuning jobs are often long running and are not attached to a worksheet session.”
      ↩︎ Cortex Fine-tuning: adapting a base model to your task
      “running the AI_COMPLETE function on a fine-tuned model incurs compute costs based on the number of tokens processed”
      ↩︎ Cortex Fine-tuning: adapting a base model to your task
      “In the navigation menu, select AI & ML » AI Studio.”
      ↩︎ Ways to reach Cortex: AI & ML Studio, SQL, Python and REST
      “If your table or view does not contain columns with the required names, use a column alias in your query to name them.”
      ↩︎ Exam trap 4
      “Cortex Fine-tuning is a fully managed service that lets you fine-tune popular LLMs using your data, all within Snowflake.”
      ↩︎ Checkpoint
    8. 8.
      “You can deploy any model from the Snowflake Model Registry as a managed service with a dedicated HTTP endpoint.”
      ↩︎ Bringing your own model: Model Registry and Snowpark Container Services
      “Snowflake simplifies the deployment pipeline by hosting your model as an HTTP server within Snowpark Container Services (SPCS).”
      ↩︎ Bringing your own model: Model Registry and Snowpark Container Services
      “Run large-scale models on distributed GPU clusters for high-performance requirements.”
      ↩︎ Bringing your own model: Model Registry and Snowpark Container Services
    9. 9.
      “CoCo is delivered through three experiences: in Snowsight, as a standalone desktop IDE, and as a command line interface (CLI)”
      ↩︎ Coding assistants: Cortex Code and Snowflake Copilot
      “CoCo knows which SQL file or notebook you are currently viewing and uses that as background context for its answers.”
      ↩︎ Coding assistants: Cortex Code and Snowflake Copilot
      “The CLI can invoke local bash commands, run git operations, and execute SQL directly against your Snowflake warehouse.”
      ↩︎ Coding assistants: Cortex Code and Snowflake Copilot
    10. 10.
      “Snowflake Copilot is being replaced with Cortex Code.”
      ↩︎ Coding assistants: Cortex Code and Snowflake Copilot
      “Snowflake Copilot is powered by a model fine-tuned by Snowflake that runs securely inside Snowflake Cortex”
      ↩︎ Coding assistants: Cortex Code and Snowflake Copilot
      “To use Snowflake Copilot in other regions, set the CORTEX_ENABLED_CROSS_REGION parameter.”
      ↩︎ Coding assistants: Cortex Code and Snowflake Copilot

    Ready to test yourself?

    Practise the 31 questions on this subdomain.

    Spotted a mistake, or was something unclear? Tell us.