Question

A Generative AI Engineer has built a RAG chain in a Databricks notebook using LangChain LCEL. The chain composes a Databricks Vector Search retriever, a ChatDatabricks foundation model, and a prompt template. The team wants to register this chain in Unity Catalog and deploy it to a Model Serving endpoint so it can be queried via REST. Which approach correctly packages the chain while preserving its dependencies (retriever config, model endpoint, prompt) for reproducible serving?

Accepted Answer

Use mlflow.langchain.log_model() to log the chain, specifying the model as code or the chain object along with required pip dependencies, then register it to Unity Catalog and create a serving endpoint from the registered model version.

Answer

Pickle the LCEL chain object with Python's pickle module, upload the .pkl file to a Unity Catalog Volume, and point the serving endpoint directly at the Volume path.

Answer

Export each component (retriever, model, prompt) as separate JSON config files and manually reconstruct the chain inside a custom endpoint init script at deploy time.

Answer

Save only the prompt template text to Unity Catalog and rely on the serving endpoint to auto-discover the Vector Search index and foundation model at runtime.