GitHub - KranzL/Jevflake: Ask TypeSafe's Jev decision model questions about your data from inside Snowflake. dbt package plus a Terraform module.

Jevflake is a dbt package that integrates Snowflake with Jev, a decision model from TypeSafe AI. Instead of generating text, Jev takes a row of data and a typed question, then returns a typed answer with a probability — making it well-suited for SQL operations where you can filter, join, and test the results. The package handles three setup tasks: configuring Snowflake to call the Jev API, creating four SQL functions (jev_noul for yes/no questions, jev_choice for picking from options, jev_score for rating against ordered levels, and jev_ask for batching multiple questions), and providing dbt macros to store and reuse answers. A Terraform module exists for teams managing Snowflake that way, though this project isn't officially affiliated with TypeSafe AI, Snowflake, or dbt Labs. Getting started requires a Snowflake account with external access enabled (not default on trial accounts), a role that can create integrations, a TypeSafe API key, and dbt 1.8 or newer. You create a secret to store the API key, add the package to your dbt_packages.yml, then run a setup operation that creates a network rule restricting traffic to api.typesafe.ai, an external access integration, the SQL functions, and grants permissions to specified roles. The setup can be split between an admin and a dbt role, and you can run it multiple times if settings change. The three function types return different structures — noul gives a 0-to-1 probability, choice returns the pick with per-option probabilities and confidence, and score returns a weighted average score with probabilities and confidence. The recommended approach stores answers in an incremental table using the jevflake.judgments macro, which only re-sends rows to Jev if their content changes or you modify the questions. You define which columns Jev reads (only what's needed for accuracy), specify a unique key like ticket_id, and map out your questions with their criteria. Individual SQL functions let you query Jev directly in select statements, while jev_ask batches multiple questions into one API call per row, cutting costs and latency when you need several answers about the same data.