Overview

Important: Snowflake CoCo is powered by LLMs and is non-deterministic. The code it generates may differ from what is shown in this guide. Always review the output and verify that the results match your expectations before proceeding to the next step.

If you prefer Snowsight Workspaces with interactive visualizations and Warehouse-based batch inference, see the companion guide: Getting Started with Snowflake CoCo in Snowsight for Data Science ML.

What You Will Learn

What You Will Build

Prerequisites

Setup

curl -LsS https://ai.snowflake.com/static/cc-scripts/install.sh | sh

What role am I using and what databases can I access?

USE ROLE ACCOUNTADMIN;

CREATE DATABASE IF NOT EXISTS CORTEX_CODE_ML_DB;
CREATE SCHEMA IF NOT EXISTS CORTEX_CODE_ML_DB.CORTEX_CODE_ML_SCHEMA;
CREATE WAREHOUSE IF NOT EXISTS CORTEX_CODE_ML_WH
  WAREHOUSE_SIZE = 'XSMALL'
  AUTO_SUSPEND = 60
  AUTO_RESUME = TRUE;

USE DATABASE CORTEX_CODE_ML_DB;
USE SCHEMA CORTEX_CODE_ML_SCHEMA;
USE WAREHOUSE CORTEX_CODE_ML_WH;

CREATE OR REPLACE FILE FORMAT ml_csvformat
  SKIP_HEADER = 1
  FIELD_OPTIONALLY_ENCLOSED_BY = '"'
  TYPE = 'CSV';

CREATE OR REPLACE STAGE ml_ltv_data_stage
  FILE_FORMAT = ml_csvformat
  URL = 's3://sfquickstarts/sfguide_getting_started_with_cortex_code_for_ds_ml/ltv_transactions/';

CREATE OR REPLACE TABLE ML_LTV_TRANSACTIONS (
	CUSTOMER_ID VARCHAR(16777216),
	TRANSACTION_TIME TIMESTAMP_NTZ(9),
	AMOUNT NUMBER(12,2),
	PRODUCT_CATEGORY VARCHAR(15),
	CHANNEL VARCHAR(8)
);

COPY INTO ML_LTV_TRANSACTIONS
  FROM @ml_ltv_data_stage;

SELECT 'Setup complete — ML_LTV_TRANSACTIONS loaded.' AS STATUS;

Optional - Generate Synthetic Data

Generate realistic looking synthetic e-commerce transaction data in
CORTEX_CODE_ML_DB.CORTEX_CODE_ML_SCHEMA. Create a table ML_LTV_TRANSACTIONS
with ~100000 transactions from ~500 customers over an 18-month period. Include
CUSTOMER_ID, TRANSACTION_TIME, AMOUNT, PRODUCT_CATEGORY, and CHANNEL. Make the
data realistic: customers should have varying purchase frequencies (some buy
weekly, others monthly), amounts should vary by category (electronics $50-$2000,
groceries $10-$200, apparel $20-$500), and channels should be web, mobile, or
in-store. About 10% of customers should be high-value (frequent buyers with
higher average spend).

Replace CORTEX_CODE_ML_DB.CORTEX_CODE_ML_SCHEMA with your target database and schema if you have one.

Confirm Sample Data

Confirm that #CORTEX_CODE_ML_DB.CORTEX_CODE_ML_SCHEMA.ML_LTV_TRANSACTIONS exists and that it contains data. Show me:
- total row count and number of unique customers
- date range of transactions
- average transaction amount by product category
- a few sample rows

Tip: The # prefix tells Snowflake CoCo CLI to resolve the table reference directly, grounding the conversation to that specific object.

Exploratory Data Analysis (EDA)

#CORTEX_CODE_ML_DB.CORTEX_CODE_ML_SCHEMA.ML_LTV_TRANSACTIONS Do EDA and recommend the features needed to train a regression model that can predict each customer's total spend in the next 90 days

Tip: Run Snowflake CoCo CLI inside a VS Code or Cursor terminal to view generated Notebook files side-by-side.

Train the Model

#CORTEX_CODE_ML_DB.CORTEX_CODE_ML_SCHEMA.ML_LTV_TRANSACTIONS
Build those features and train a regression model to predict each customer's total spend in the next 90 days. Use two different algorithms and evaluate the best one. Use 20% of the data as the eval set.

Log to Model Registry

#CORTEX_CODE_ML_DB.CORTEX_CODE_ML_SCHEMA.ML_LTV_TRANSACTIONS
Log the better model with metrics into Snowflake Model Registry, and use SPCS to create a REST endpoint for online inference.

Run Online Inference

Create feature profiles for 50 customers and run LTV predictions using the REST
API for online inference running on SPCS. Show the top 10 highest predicted LTV
customers and a latency profile (p50, p95, p99).

Conclusion and Resources

What You Learned

Related Resources