[Chore]: Refine Self-Correction Loop to Prevent Prompt Bloating

[chiruu12]

Description:

Context:
This issue is in response to feedback received on PR #7 regarding our self-correction mechanism:

"The feedback can cause bloating. I don't exactly know how much token each iteration adds, but I feel the original problem gets buried under feedback noise!"

Problem:
The current implementation of our self-correction loop appends the critic's feedback directly to the end of the original prompt for each retry attempt. As noted in the feedback, this approach has two main drawbacks:

1. Prompt Bloating: It progressively increases the token count with each iteration, leading to higher costs and latency.
2. Context Dilution: The original, core problem description gets pushed further up in the context window, potentially reducing its significance to the LLM and "burying it under feedback noise."

[Winstar444]

assign

[Winstar444]

/assign

[Winstar444]

Hi @chiruu12 ,

I’d like to work on this issue and start contributing to the project.
I’m currently setting up the project locally and reviewing the related code paths.

Thanks for your time!

[chiruu12]

@Winstar444 2 people are already working on it. How about finding any other issue?