What’s the recommended architecture for integrating Databricks + MLflow + GenAI for model training

[Suheb]

If I want to use Databricks, MLflow, and GenAI together, what is the best way to organize and connect them so that I can train AI models and then use them in real apps?

[Raman_Unifeye]

Broad question. I will recommend to follow Databricks offical documents and ML Training at their customer portal.

Yet, I will try to answer it as below

Use the Databricks Lakehouse as the unified platform, with Unity Catalog (UC) providing centralized governance for all data and models. First, store your raw and processed data in UC-governed Delta Tables, using Mosaic AI Vector Search to create retrieval indices for RAG applications. Next, leverage MLflow Tracking to monitor your model development; the critical part for GenAI is using MLflow Tracing to log every step, tool call, latency, and cost of your LLM pipelines. After validating your model with LLM Judges using MLflow Evaluation, register the best version in the UC-governed MLflow Model Registry. Finally, deploy the model via a Databricks Model Serving endpoint, which provides a scalable API for your external apps and automatically logs production traces back to the Lakehouse, closing the continuous improvement loop.

This video demonstrates how MLflow 3.0 provides a unified platform for AI and MLOps, essential for GenAI. MLflow 3.0: AI and MLOps on Databricks - YouTube