[native_doc_dartifier] Last output in retry loop not always the best

[marshelino-maged]

When feeding the analysis error again to the LLM in a loop to make it solve the problems ... sometimes it gets a worse response or makes some changes that result in more other analysis errors ... currently we output the last response as the result ... but instead we can output the best one of them ... might be the one with the least analysis errors.

[dcharkes]

Or, if it goes down the wrong path, maybe we should go back to the result with the least analysis errors and restart it from there? Or is it too deterministic and it will take the same wrong path?

[marshelino-maged]

Or, if it goes down the wrong path, maybe we should go back to the result with the least analysis errors and restart it from there? Or is it too deterministic and it will take the same wrong path?

With high probability, it will take the same wrong path if retry again with the same previous context.
But we could give it the wrong path it takes as a history ... and prompt it again from the least analysis errors point to try to solve those errors in a different way.