Great walkthrough, @BijuThottathil!

I really liked how you tied the full flow together: curated Delta table, Vector Search, MCP endpoint, and Claude Desktop via a local bridge. That end-to-end view is what’s usually missing in most demos.

Two points that stood out to me:

1. Building the search_text column from structured attributes to create a richer embedding signal is a very practical pattern. It is often overlooked but makes a big difference for retrieval quality.

2. Using a local MCP bridge to authenticate and call the Databricks Vector Search tool cleanly separates concerns and keeps things flexible.

One question: have you tested how this behaves with incremental updates to the source Delta table? I am curious how you would approach re-embedding or managing latency as the table evolves. 🙂

Good Question Edonaire. There 2 options here. Continuous and Triggered. If continuous, it will take care of incrimental load automatically, but costly.

Triggered can be filed by setting up job with a task to run the "Refresh" code.