Versioning prompts and managing their lifecycle

Databricks uses MLflow's Prompt Registry (built on Unity Catalog) to version, govern, and promote prompt templates independently of application code, mirroring how MLflow Model Registry manages model versions. This lets engineers track prompt history, compare versions, and control which version is live in production without redeploying the app.

1 · Learn the must-know

• 1Prompts are registered as immutable versioned objects (e.g., via mlflow.genai.register_prompt) in Unity Catalog, so each edit creates a new version rather than overwriting the old one.
• 2Prompt aliases (such as 'production' or 'staging') can be reassigned to point at different versions, enabling promotion/rollback of a prompt without changing application code.
• 3Prompt templates use double-curly-brace variables (e.g., {{question}}) that are filled in at inference time, keeping the template reusable across contexts.
• 4The MLflow UI/API lets you view full prompt version history, diff two versions, and see lineage linking a prompt version to the runs or model versions that used it.
• 5Loading a prompt in code by alias (not hardcoded text) decouples app logic from prompt content, so prompt updates don't require code redeploys.
• 6Prompt registration and lifecycle management works alongside standard MLflow experiment tracking, so prompt changes can be correlated with evaluation metrics to decide promotion.

2 · Check your understanding