Hi, for me, I will try my best to answer. In my experience, it all about giving the agent the right tool for job. We give them a grep tool and search, then the accuracy and reliability rely on the capable of the running model.
> I ask the agent to generate things like:
> { original_sentence: xxxx, improved_sentence: yyyy }
> and I use the original_sentence to locate the original content, and replace it with the improved_sentence.
To mitigate this, I build the apply patch pattern. [0][1]
> I ask the agent to generate things like: > { original_sentence: xxxx, improved_sentence: yyyy } > and I use the original_sentence to locate the original content, and replace it with the improved_sentence.
To mitigate this, I build the apply patch pattern. [0][1]
[0] https://github.com/vinhnx/vtcode/blob/main/vtcode-core/tests...
[1] https://github.com/vinhnx/vtcode/blob/main/vtcode-core/src/t...