Running Open-Jev with SQL on Databricks

Author: Cyrus Cowley, Tejas Sundaresan, Hanlin Tang, Anshul Gupta, Julia Powell (Databricks)

In short: Open-source decision models can now run on a managed platform, callable directly via SQL, without your data ever leaving your governed environment.

I. Why This Matters

The open-source decision model space has been buzzing lately—from SemIf-OpenJev to Laya to Kev, there’s been a steady stream of open-weight releases.

But these models only unlock their true power when they can access large datasets.

And “accessing large datasets” is precisely what data platforms are built for. So the path laid out by Databricks connects three key components:

• Open-source decision models (they’re using SemIf-OpenJev)
• Governed enterprise data
• SQL, the interface everyone already knows

II. Three Steps to Start

They’ve packaged an importable Notebook. Getting started takes just three steps:

1. Select the model name/schema
2. Choose Serverless GPU
3. Click “Run All”

This Notebook uses Databricks AI Runtime to provide serverless, on-demand GPU compute—no need to set up or manage GPU infrastructure yourself.

It performs three tasks:

• Downloads the model files for SemIf-OpenJev
• Logs and registers the model using Express Deployments
• Automatically creates a GPU Model Serving endpoint

At this point, you have an open-source classifier running behind a managed endpoint, ready to be called from applications or directly via SQL.

III. Key Insight: The Endpoint Doesn’t Need to Expose a Standard Chat Interface

This is, in my view, the most important sentence in the entire piece:

A useful aspect of this setup is that the endpoint does not need to expose a standard chat-completions interface.

Databricks’ ai_query supports arbitrary custom model APIs. This means you can send SemIf’s structured request format directly from SQL.

This implies you don’t have to alter the model’s interface to accommodate the platform. Decision model interfaces were never chat-based, and this platform accepts that.

The result: you can apply the model to evaluation data already residing in Databricks, and retrieve the selected classifications and option probabilities as structured columns.

IV. My Assessment

The value here isn’t just “Databricks launched a feature”—it’s the pattern it demonstrates.

Previous articles we wrote on using decision models mostly stopped at the “how to make the model give a judgment” stage—how to construct prompts, read logprobs, normalize probabilities.

What this adds is the crucial next step: once the judgment is made, how do you use it?

The answer it provides is “turn it into a SQL function that can run on an entire table.” This form transforms a decision model from “a service you call” into “an operator you write into your query.”

Three concrete benefits:

1. Data stays within the platform. The model runs alongside your data, rather than sending data out and fetching results back—this is decisive for enterprises with compliance requirements.
2. Naturally suited for batch scenarios. Decision models are cheap, fast, and deterministic (a property we’ve repeatedly highlighted). These three traits together are exactly what’s needed to “run the same judgment on millions of rows.”
3. Replaceable. Because the endpoint accepts any custom API, swapping SemIf-OpenJev for another open-source decision model requires no changes to the SQL.

One boundary to note: this is a vendor blog, primarily guiding you to use their platform. The Notebook and endpoint creation steps are tied to Databricks.

However, the pattern it demonstrates is transferable: if your decision model can be called as a SQL function, it can integrate into existing data pipelines, rather than being siloed in another service waiting to be integrated.

Links:

• Original: https://www.databricks.com/blog/running-open-jev-sql-databricks
• SemIf-OpenJev Repository: https://github.com/TheoLeeCJ/SemIf-OpenJev
• Databricks ai_query Documentation: https://docs.databricks.com/aws/en/sql/language-manual/functions/ai_query
• Companion Notebook: https://docs.databricks.com/aws/en/notebooks/source/machine-learning/serve-semif-decision-models.html

#Jev #Databricks #SQL