feat(docs): RFC on Searching for Focus Component on UI

[LaPetiteSouris]

What kind of change does this PR introduce?

An RFC in attempt to solve #421 and to provide a direction for #442

Summary

When working on #421, I realize that there are a few missing pieces of the puzzle in an attempt to evaluate a prediction of the models. Mainly, we do not have yet the definition of "how good and useful" a prediction is. Without this, it is hard to evaluate and see if a model provides correct prediction or not.

Somehow, this problem is entangled with the fact that we need to come up with a prompt strategy and a strategy to summarize a window data into useful hints for the models.

This PR provides a baseline for discussion on how we can decompose the windows into useful elements which later can serve not only model evaluation process, but also potentially prompt building as well as to some extend, RFLHF process.

Checklist

• My code follows the style guidelines of OpenAdapt
• I have performed a self-review of my code
• If applicable, I have added tests to prove my fix is functional/effective
• I have linted my code locally prior to submission
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation (e.g. README.md, requirements.txt)
• New and existing unit tests pass locally with my changes

How can your code be run and tested?

The core RFC file should be reviewed/discussed. No code is written yet at the moment

Other information

• Not sure where to put the RFC file, hence I created it in openadapt.docs.
• It is more convenient to just read the RFC file

[abrichr]

Thank you @LaPetiteSouris ! As per our conversation last week, this sounds directionally correct. Before implementing, can you please coordinate with @FFFiend to parallelize the work? 🙏

The first step of the problem formulation is:

1. At a single time step, given an action and a tree, find the target in the tree

@FFFiend can you please suggest some next steps once this has been implemented?

[LaPetiteSouris]

Thanks

With pleasure. This week I'll come up with a kind of skeleton code proposals to make sure the this is compatible with future fine-tuning, completion provider. If @FFFiend has time, please be one of the reviewers for that coming code/skeleton proposal for the next steps.

@FFFiend how far are you off to finish FineTuning. Can I wrap up another round of review ?

I would love to review CompletionProvider as well when it is available.

[FFFiend]

Thanks

With pleasure. This week I'll come up with a kind of skeleton code proposals to make sure the this is compatible with future fine-tuning, completion provider. If @FFFiend has time, please be one of the reviewers for that coming code/skeleton proposal for the next steps.

@FFFiend how far are you off to finish FineTuning. Can I wrap up another round of review ?

I would love to review CompletionProvider as well when it is available.

#379 is ready for review. All done. Fine-tuning (#453 is too, however based on the discussion here and our steps moving forward, I think that PR is an example of a failure mode of fine-tuning, i.e showing the performance of the Davinci GPT model on the basic bare-bones iteration of Action, Window pairs)