We can likely make prompt and model iterations and experimentations easier with the mms server. Currently it requires restarting the server, and running the accuracy tests from Compass. This ticket involves investigating ways we can make the developer process here smoother and faster for validating changes.

We should enumerate the pain points as well so we know what solutions will achieve. Server restart times, model interchangeability requirements, etc.

While this investigation is open ended, two suggestions on things to possibly investigate are:

• A minimal java server for local development. Use the code from the prompt utils in mms and performs the request handling, without all of the other parts of mms. It would need to have the model interchangeable, and spinup/teardown quick.
• Move the accuracy tests to java so they can be run as unit tests without requiring the server restart.