Sourcegraph Cody is Generally Available (Free and $9/month)

I'm not associated with this project, just very excited about it!

> Cody's code completions differ from standard AI code completions through the use of an open source LLM enhanced with Cody's context engine. Specifically, we've found the type of context that helps a lot with completions is graph context. Graph context takes advantage of Sourcegraph’s expertise in parsing code and producing a graph that captures semantic understanding of your code. Pulling in this semantic code graph into context reduces the rate of common LLM hallucinations like type errors and imaginary function names.

> By using the open source StarCoder LLM, we've been able to make completion latency very low. It also gives us access to signals such as token-level log probabilities, which we can incorporate into end-user signals.

> The combination of these qualities allows us to hit a Completion Acceptance Rate of 30% or higher, depending on the scenario. And we’re just getting started. Upcoming versions of Cody will integrate more of Sourcegraph's universal code graph into the context engine. Users of Cody will see these improvements in the coming weeks and months.

I made the move from Copilot to Cody about a month ago. I've noticed I'm accepting more of it's suggestions. I was pretty sure it tracked state on variables and functions because it's suggestions seem to flow from one file to the next. Cool to see this post explaining that magic. Pretty nifty!

Hm, what about privacy? Do they store code on their servers?