Intelligent Cody: MVP using VSCode API

[efritz]

This is an experimental (currently internal) MVP of fetching first-order definitions of symbols available in the current file. See the Intelligent Cody update for additional context.

This PR supplements the context fetching mechanism used to grab embeddings to also utilize the local VSCode API to emulate what we've been doing in the server-side implementation https://github.com/sourcegraph/sourcegraph/pull/51690.

Note: much of this code is shared from Cesar's work in #78, and my hackathon project in #472.

Test plan

Tested locally, and gated all changed code with an experimental VSCode extension setting feature flag.

[beyang]

Can you provide some screenshots and/or illustrative examples of where this improves response quality and where it hurts response quality?

Is it targeted at autocomplete, chat, or all of the above?

[efritz]

Can you provide some screenshots and/or illustrative examples of where this improves response quality and where it hurts response quality?

Sure, will gather some and post in a comment on the PR.

Is it targeted at autocomplete, chat, or all of the above?

Right now it's only set up for chat, but I'm currently working (between a few other tasks) to thread the graph context fetcher through autocomplete as well.

[toolmantim]

For the name, I'm trying to move to a single bit of UI that summarises all the context prep Cody did and was included (vs the separate drop-downs we're adding here). The name would sit alongside "Embeddings".

My ranked naming suggestions (instead of "Intelligent Cody"… Cody is already intelligent in lots of ways 😉)

1. Add experimental support for locally generated Symbol Context
2. Add experimental support for locally generated Precise Symbol Context
3. Add experimental support for locally generated Code Intel Context

imo from a product design POV calling it "Symbol Context" and using the word "Symbols" is probably best, alongside "Embeddings". And we can have other single-word bits of context, like "Ownership"and what-have-you.

[valerybugakov]

That's awesome! Since all the changes are behind the experimental flag, we can merge this one safely and iterate on improvements in follow-up PRs.

[valerybugakov]

Is this structure based on the work that's happening on the server-side? (should we keep in sync with something?)

[efritz]

It's meant to conform to the GraphQL schema we've defined here, so keeping it in the same shape will make the Remote API integration easier since we can just return the same type directly.

[valerybugakov]

Would it be helpful to leave a comment on that?

[valerybugakov]

Also, we should add an entry about the new experimental setting to vscode/CHANGELOG.md.

[efritz]

@beyang Here's a quick example. I highlighted the part of the code I wanted precise symbols to be extracted from, then asked something about the shape of the type of one of the struct members (defined in the stdlib).

Clicking on the sidebar jumps me to ~/go/pkg/mod/golang.org/x/crypto@v0.11.0/ssh/certs.go:

[beyang]

For the name, I'm trying to move to a single bit of UI that summarises all the context prep Cody did and was included (vs the separate drop-downs we're adding here). The name would sit alongside "Embeddings".

@toolmantim we should not use "embeddings" anywhere that's visible to the end user. Embeddings are an implementation detail and we need another descriptor for it, like "AI-enabled search" or "Intelligent search".

[efritz]

How do I turn it on to try it out?

There's an extension setting, mentioned in the changelog. It's not reactive so you'll need to reload the extension once you enable it (for now).