Snowflake Cortex AI Features: AI Functions, Fine-tuning, Cortex Search and Cortex Analyst

1.The principle: Gen AI that runs where the data lives

Snowflake Cortex is the name for Snowflake's suite of AI features built on large language models. It includes AI Functions, Cortex Fine-tuning, Cortex Search, Cortex Analyst, Cortex Agents and Cortex Code. One principle sits under all of them: the models come to the data. Every LLM that Snowflake gives you access to is deployed inside the Snowflake Service perimeter. Your prompts and table contents therefore never have to be shipped to a third-party endpoint that you manage yourself.

The second principle is about training. The Cortex Analyst documentation states the policy directly: Snowflake does not use your Customer Data to train or fine-tune any model that is made available across its customer base. By default, Cortex Analyst runs on Snowflake-hosted LLMs from Mistral and Meta, so no data, metadata or prompts leave the governance boundary. The third principle is that existing access control still applies. Cortex Analyst integrates fully with role-based access control (RBAC), so the SQL it generates and runs obeys the same grants as any other query. Calling an AI feature is itself a privilege: AI Functions need the USE AI FUNCTIONS account-level privilege plus the CORTEX_USER or AI_FUNCTIONS_USER database role.

Sources12

2.Cortex AI Functions: LLMs as SQL

The most direct way to use Cortex is through AI Functions. These are managed functions that you call from SQL, and they are also available in Python. They run LLMs from OpenAI, Anthropic, Meta, Mistral AI, DeepSeek and xAI over text, images and documents. Most are task-specific and need no customisation. AI_COMPLETE is the general-purpose function for most generative tasks.

AI Functions are tuned for throughput. They work best when you run them over large SQL tables in batch. When latency matters, as in an interactive app, Snowflake points you to the REST API. It offers the Complete API for simple inference, the Embed API for embeddings, and the Agents API for agentic applications.

Sources3

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

Sometimes prompt engineering, or even RAG, cannot give you the quality or latency a specialised task needs, and training a large model from scratch costs too much. Cortex Fine-tuning covers that middle ground. It is a fully managed service that uses parameter-efficient fine-tuning (PEFT) to build customised adaptors on top of a pre-trained model, using your own examples and without the data leaving Snowflake. The documented base model is llama3.1-8b, which has a 24K context window. Snowflake notes that the list of models may change.

Fine-tuning is exposed as one Cortex function, FINETUNE, which takes four arguments: CREATE, SHOW, DESCRIBE and CANCEL. Jobs often run for a long time and are not tied to a worksheet session. You pay for trained tokens, which are input tokens multiplied by epochs. Running AI_COMPLETE against the fine-tuned model is charged separately, and standard storage applies to the adaptors. The role that creates the job needs CREATE MODEL (or OWNERSHIP) on the target schema.

SELECT *
  FROM SNOWFLAKE.ACCOUNT_USAGE.CORTEX_FINE_TUNING_USAGE_HISTORY;

Sources4

4.Cortex Search: retrieval for RAG and unstructured data

A fine-tuned or general-purpose LLM still knows nothing about the document you added to a table this morning. Cortex Search closes that gap. It is a managed hybrid search engine that combines vector and keyword search over your text data. Snowflake handles the embeddings, the infrastructure, quality tuning and index refreshes for you. Its two primary use cases are retrieval augmented generation (RAG) and enterprise search. In RAG, Cortex Search retrieves relevant passages and an LLM function such as AI_COMPLETE writes an answer grounded in them. Cortex Search is also the retrieval layer for Cortex Agents when they work over unstructured data.

Unstructured data reaches Cortex Search in two ways. You can index text columns in a table or view, for example text produced by AI_PARSE_DOCUMENT, which the docs describe as suited to parsing documents for analytics and RAG pipelines. In Snowsight you can also point a service directly at files in a stage, a capability marked Preview. A service is one SQL statement. TARGET_LAG controls how far the index may lag behind the base table.

CREATE OR REPLACE CORTEX SEARCH SERVICE transcript_search_service
  ON transcript_text
  ATTRIBUTES region
  WAREHOUSE = cortex_search_wh
   ?  = '1 day'
  EMBEDDING_MODEL = 'snowflake-arctic-embed-l-v2.0'
  AS (
    SELECT
        transcript_text,
        region,
        agent_id
    FROM support_transcripts
);

Sources53

5.Cortex Analyst: text-to-SQL over structured data

Cortex Search is for unstructured text. Cortex Analyst is its counterpart for structured data. It is a fully managed, LLM-powered agentic system that turns a business user's natural-language question into SQL, then runs that SQL in your own virtual warehouse. The model sees only metadata (table and column names, types and descriptions), and RBAC applies to the query it produces. Accuracy comes from semantic views. These are schema-level objects that define logical tables, dimensions, facts, metrics and relationships, and they can include verified example questions. Snowflake recommends them over the older YAML semantic model files on a stage. Cortex Analyst is API-first: it is delivered as a REST API so you can embed it in Streamlit apps, Slack, Teams or a custom chat interface.

Sources1

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.AI Functions such as AI_COMPLETE are the best choice for low-latency, interactive applications.Why is that wrong?

   AI Functions are optimised for throughput and batch processing over tables. For interactive, latency-sensitive use, Snowflake recommends the REST API (Complete, Embed or Agents API).

   Covered in Cortex AI Functions: LLMs as SQL

2. 2.Cortex Fine-tuning trains a new large model from scratch on your data.Why is that wrong?

   It uses parameter-efficient fine-tuning (PEFT) to build adaptors on top of an existing pre-trained base model.

   Covered in Cortex Fine-tuning: adapting a base model to your task

3. 3.Cortex Analyst only needs a database schema to generate accurate SQL.Why is that wrong?

   Schemas lack business definitions. Cortex Analyst relies on a semantic view (or legacy semantic model YAML) to describe metrics, dimensions and joins.

   Covered in Cortex Analyst: text-to-SQL over structured data