???

Aim

Experiment with retrieval augmented generation in scala

Why?

I’ve been on the “business side” of some of the proposed solutions, and they didn’t fulfil the needs. In particular, the challenge of

  • generating prompts
  • gaining an intuition for useful promting
  • ensuring future quality

Caused me quite some concern.

Status:

Conceptual

Prior Art

https://github.com/explodinggradients/ragas

Notes

I think there are three key ways we could consider differenting this project.

  1. Mills fine grained caching, could allow a constant time feedback loop, even as number of documents and prompts grow.
  2. If we write a X platform API, it shoudl be possible to expose diagnostic information to the user via a web api. This could be used to build a UI to help users understand the quality of their prompts.
  3. Consider seperation of retrieval via cosine search, and generation.