bars0um
commented
5 months ago So far this is proving harder than I thought. I've removed a lot of technology specific wordage but some models swerve into a totally unexpected set of technologies after getting it right for a couple of conversations. I am not sure if this is sourced in the models themselves or if there are remnants within the code itself or the prompts that are somehow pushing the conversation in that direction. I persistently for example find that it chooses to go down a node path instead of using views and javascript for the frontend. Needs more exploration, I suspect also that the amount of instructions that are in the prompts are possibly flooding the attention window such that some of the original guiding text is lost...I see that in lots of steps, a re-hashing of the original plan is inserted but perhaps more of that needs to happen...exploring further
blarvin
Version
Command-line (Python) version
Suggestion
I'm trying to see if I can add support for rails. I'm running into a lot of situations where the LLM diverges totally because there are minor references sprinkled around the code to specific technologies like Node.js, React, Redux, etc.
It would be best if these were all templatized so that you could easily and seamlessly add support for other languages. Once I have a working copy for rails, I'm happy to provide a repo containing all the references I had to change to get it to work. We could then open a PR to convert these to templatized variables.