Code completion

[cathaysia]

Will this plugin add code completion? Just like this: https://github.com/monkoose/neocodeium

[olimorris]

Probably not. It's something I have thought about greatly. But I doubt I'll be able to come close to what Copilot.vim or Copilot.lua do. Also, for anything other than ollama models, it's too expensive.

[cathaysia]

In my personal experience, Ollama code generation is slower than Copilot. What do you mean of expensive, do you mean that the requests to the server are too frequent?

[cathaysia]

If we increase the debounce time, may be able to reduce the cost?

[olimorris]

Expensive in terms of token cost. Have you found any models that can be connected to Ollama that do suggestions well?

My other reservation is that the LLM will need significant context about the workspace and the user's most recent edits, in order to determine the suggestion, which would take some time.

[cathaysia]

As of now, I haven't found a good code complete plugin used Ollama. The best chat plugin experience is this plugin.

Currently, codellama:7b is fast, with a delay of only about one second.

[dlants]

I think it's worth reconsidering. It could be disabled by default... It also could be hooked up to be manually prompted... like "finish what I'm trying to say here", but with a more immediate & polished UX than :cc ? Maybe there's already a way to hook codecompanion up to "finish what I'm trying to do here?

There's also an expanding set of things that this could lead to...

I think windsurf keeps track of all the edits in the file so far, and puts those in the context when sending to their model. They also seem to ask for edits in a wider window within the file, rather than just a continuation past the cursor, the result is a pretty slick experience that can often guess what you're intending to do next.

I'll try using this side-by-side with a manual copilot inline complete and report back :)

[olimorris]

If I change my mind on this I will let you all know. As I've stated before:

1. Copilot.lua is a fantastic plugin that likely won't be beaten by this. So I'm loathed to spend weeks implementing something which I won't use.

2. The workspace context needs to come first (I'm slowly working on it) and will need a lot of trial and error and continued output being sent to the LLM...

3. Which is expensive. Sure, users can turn it off but then I've created a feature that took weeks which very few users will leverage...

4. Because Ollama models are currently still way below par for most average laptop users.

[PLAZMAMA]

FYI, if you want code completion with local LLMs, checkout Tabby