Train a model from a Snowflake or Databricks table, and fill new rows with its answers

• An Essentials or Business workspace. The Owner connects the warehouse; Builders choose a listed table to train or fill from.
• Snowflake: a role that can create a user and a role and grant on the tables, such as ACCOUNTADMIN, and a virtual warehouse for StayCharted's queries.
• Databricks: a workspace with Unity Catalog, a SQL warehouse (serverless or pro), and permission to create a service principal and grant on the catalogs.
• The queries run on your compute, so Snowflake or Databricks bills them as usual.

In StayCharted, open Settings → Warehouses and choose Snowflake or Databricks.

Press Connect and show the grant SQL.

StayCharted shows the SQL that gives it exactly what you listed, and nothing else. Copy it and run it in your warehouse.

Snowflake (a Snowsight worksheet, as ACCOUNTADMIN or a role that can create users and roles):

• creates the results schema if it isn't there;
• creates a role and a service user (TYPE = SERVICE) for StayCharted, and sets the public half of the key StayCharted made. The private key stays in StayCharted, encrypted, and is never shown;
• grants the role to SYSADMIN, as Snowflake recommends for custom roles, so your admins can read the results table;
• grants USAGE on the virtual warehouse, the databases and the schemas, SELECT on the tables you listed, and CREATE TABLE in the results schema.

Databricks (the SQL editor, as an admin of these catalogs):

• creates the results schema if it isn't there;
• grants USE CATALOG and USE SCHEMA, SELECT on the tables you listed, and CREATE TABLE in the results schema, to the service principal.

Also give the service principal Can use on the SQL warehouse, from the warehouse's Permissions button.

Press Test. StayCharted signs in, reads the columns of each table you listed, and creates its results table. Each line says what worked, or the warehouse's own words when something didn't, such as a missing grant.

Start a new model and choose Or read the rows from a Snowflake or Databricks table. Pick the table, tick the columns the model should read, and choose the column with the answers your team gave. StayCharted shows the table's columns and a few rows first.

Rows with no answer are left out. The rows become the model's training file, exactly as an uploaded spreadsheet would. From there it's the usual steps: check the data, train, and read the report on rows the model never saw.

On the model's Use tab, under Fill a spreadsheet, choose From a warehouse table.

1. Pick the table, and the column that identifies each row (its ID).
2. Leave Only rows this model hasn't answered yet ticked. StayCharted then reads only rows with no answer from this model in the results table, so a weekly fill picks up only what was added since.
3. Press Read the rows, then Run and fill in categories.

The model reads the same columns it was trained on, matched by name. If the table names them differently, StayCharted asks which column holds each. One fill reads up to your plan's rows per file to fill; when a table has more, the next fill picks up the rest.

When the fill finishes, the answers are in your results table, and the completed file is in StayCharted too. Answers below your review cutoff are flagged NEEDS_REVIEW and wait in the Review Queue for a person.

The results table has one row per row answered and model:

A row answered again is updated in place. The model's Training files tab shows the exact SQL to join its answers back to your table. In Snowflake it looks like this:

SELECT t.*, r.CATEGORY, r.CONFIDENCE, r.NEEDS_REVIEW
FROM ANALYTICS.SUPPORT.TICKETS t
LEFT JOIN ANALYTICS.STAYCHARTED.STAYCHARTED_RESULTS r
  ON r.SOURCE_KEY = CAST(t.TICKET_ID AS VARCHAR)
 AND r.SOURCE_TABLE = 'ANALYTICS.SUPPORT.TICKETS'
 AND r.MODEL_ID = '<your model ID>';

To let another role or group read the results, grant SELECT on the results table to it.