search

gmickel / CodeWhisper

CodeWhisper: AI-Powered End-to-End Task Implementation & blazingly fast Codebase-to-LLM Context Bridge
https://discord.com/invite/82mjJkwABQ
MIT License
52 stars 4 forks source link

can it create a project from scratch? #45

Open zenitogr opened 3 months ago

zenitogr commented 3 months ago

using the npm package codewhisper in an empty folder it just terminates after displaying file and folder selection.

I havent set a AI API maybe thats the reason?

image

zenitogr commented 3 months ago

set an Env variable in powershel with $Env:GROQ_API_KEY="my groq key"

and run again, it seems that it doesnt work with an empty folder: image

gmickel commented 3 months ago

Thanks for the question @zenitogr !

CodeWhisper's main use-case is definitely implementing features, creating reviews, generating documentation, etc. on pre-existing projects. However you can easily workaround this by creating a boilerplate project before starting a codewhisper task.

For example if you wanted to build a Nextjs project, just create it first npx create-next-app@latest then do a quick git init and commit. Then you can start using CodeWhisper to add features.

zenitogr commented 3 months ago

ok how possible is it to allow it to work on a fresh empty folder? maybe a special command that it is all about initiating a project?

gmickel commented 3 months ago

If I were to add that functionality, I would probably use a second command that doesn't follow the codebase -> plan -> code generation route. I'll look into it.

Quick side note: I see you were trying to run it using the Groq API. You might have some success with it, but in my tests the rate-limits are currently far too low (if you're only a free user like I am :) for large feature implementations using CodeWhisper.

zenitogr commented 3 months ago

can I use ollama? I got a 4070 and 32 GB ram! what model should I use with ollama? can i use LM-Studio?

gmickel commented 3 months ago

Realistically, for complex implementation tasks you'll want to be using the Anthropic, OpenAI or LLama 3.1 405b models. The latter isn't yet available on Groq, but I'm looking into adding another provider for it.

But you can use Ollama. Check the quick start section or usage for examples. The prompts CodeWhisper uses are currently tuned for these very large models and I'm still the process of testing whether I can get smaller models to work using simpler prompts.

I want to test it with deepseek-coder-v2, mistral-nemo, mistral-large, etc.

I will update this and the README when I've finished evaluating.