Hacker News Reader

I think (just my experience) that copilot (the vim edition / plugin) uses more than just the current buffer as a context? It seems to improve when I open related files and starts to know function / type signatures from these buffers as well.

View on HN

That could be. If so, it would be interesting to know how Copilot does that.

For me, just asking LLMs "Can the following function be improved" for a function I just wrote is already pretty useful. The LLM often comes up with a way to make it shorter or more performant.

Yes, the official plugin sends context from recently opened other buffers. It determines what context to send by computing a jaccard similarity score locally. It uses a local 14-dimensional logistic regression model as well for some decisions about when to make a completion request, and what to include.

There are some reverse-engineering teardowns that show this.

I just tried the gpt4, without any modifications it's impressively worse than the current chat model

What did you try?

Running some queries in a new chatgpt session and via the API. I tried adding the same system prompt on both.

I can run one for you, if you want :)

"some queries"?

Show them, so we can discuss?

Reply on news.ycombinator.com