What does "Use vector encodings to automatically select code" in the roadmap mean?

[ziontee113]

Hi, I found this "Use vector encodings to automatically select code" line on the README but can't quite understand what that means. Can you please explain it to me @gptlang?

[gptlang]

It's what the Copilot Chat extension does in VSCode. It takes the file you're on and uses tree-sitter to get all functions & classes. It then submits your code to the embedding API

CopilotChat.nvim/rplugin/python3/copilot.py

Lines 136 to 150 in c54e61f

	def _get_embeddings(self, inputs: list[typings.FileExtract]):
	embeddings = []
	url = "https://api.githubcopilot.com/embeddings"
	# If we have more than 18 files, we need to split them into multiple requests
	for i in range(0, len(inputs), 18):
	if i + 18 > len(inputs):
	data = utilities.generate_embedding_request(inputs[i:])
	else:
	data = utilities.generate_embedding_request(inputs[i : i + 18])
	response = self.session.post(url, headers=self._headers(), json=data).json()
	if "data" not in response:
	raise Exception(f"Error fetching embeddings: {response}")
	for embedding in response["data"]:
	embeddings.append(embedding["embedding"])
	return embeddings

Once that's done, it finds the n most similar code snippets using cosine distance on the embedding and automatically inserts that into the prompt rather than having to manually place code snippets into the clipboard/unnamed register.

Like this:
https://github.com/acheong08/vectordb/blob/83ee790411cdeb875d1db2ee29b81bb69eb9b368/rank/rank.go#L29-L61

[ziontee113]

Couldn't we just use Neovim built-in Treesitter and LSP to do that?

[gptlang]

Yes.

[ziontee113]

I'm really not sure why embeddings are necessary. What's the use case to use embeddings anyway?

[gptlang]

There is a @workspace command that allows it to insert multiple code segments relevant automatically from across your git repository. It's quite helpful in finding code that handles something when you don't know where it is.