Building on the initial code that rh-ee-jbarea wrote.

Switched to use python-openai instead of llamastack after struggling to get llamastack working with vLLM (also, we don't need a whole stack for this)