How do you enhance LLM app response quality?

I started an open-source project that helps A/B test LLM-based apps: LangBear. It'll help you manage prompts and few-shot examples, run A/B tests, get user feedback, and see results.

As long as I cannot provide all at once, I want to know what developers need the most.

If you conduct A/B testing to enhance the quality of LLM responses, what aspect would you be most interested in testing? e.g., Prompts, Vector stores, Foundation models, or Agents