TL;DR: Host your own AI autocomplete in Visual Studio.
Only supports KoboldCpp/KoboldAI and OogaBooga, support for other APIs is not guaranteed, but you can try, if they follow the same scheme as mentioned providers.
Visual Studio plugin that adds functionality to use your own AI suggestion and autocomplete provider.
The idea came about after trying out several big language models for generating C# code. One of my concerns with AI autocompletion services always was making sure my code wasn't put at risk where others could misuse or leak it. And here is the solution to it.
Now works with IntelliCode disabled.
Features
Uses VS suggestion mechanism to provide "Tab to accept" from your local AI backend.
Adds references (Classes, Interfaces, Enums, Records, Structs) from your code to the generation prompt to give LLM a better understanding of your code (C# only).
Built-in tokenizer and KoboldCpp tokenizers to keep generation prompt from "pushing out" your code out of context when adding the references.
Hotkeys for manual generation requests and to force VS to show the suggestion again if you click away and your suggestion is gone.
Examples
Singleline
Multiline
What does L.AI stand for?
L.AI simply means Local Artificial Intelligence
Usage
Start your Visual Studio and start typing. It follows the same rules as IntelliCode.
Hotkeys:
[Alt+A] Generate code suggestion.
[Alt+Z] Generate single-line code suggestion.
[Alt+S] Show the last suggestion.
Known issues
IntelliCode loves to think that it's smart and to override suggestions from this extension from time to time. But now you can disable it completely and keep the extension's functionality.
Changelog
v1.2.1
Fixed code being trimmed on accepting the suggestion;
Fixed multiline generation issue. Multiline is now a default option;
Improved API requests to accommodate generation canceling;
Competition requests are now delayed to not overwhelm the API (configurable in settings);
v1.2.2
Added pane to display manual generations in output.
Added interactivity to keep users posted about the connection status.
v1.3.1
Added functionality to override VS suggestions completely. Now can work without IntelliCode enabled.
Added option to disable automated suggestion generation
Added a new command to generate a single-line suggestion
Reworked the Generate and Insert logic, which now generates a suggestion instead of pasting it into the document. Suggestions generated from the commands are now automatically shown as the "ghost code"
v1.5
Added an option to include reference types in the generation prompt. This only applies to commands (hotkeys).
What it means is:
When you make a manual generation request and "Use analyzer" is enabled in the options - the analyzer will lookup all references in the current file (Classes, Interfaces, Enums, Records, Structs) and append their source code at the end of the prompt, giving your LLM more context about your code, and supposedly it will improve the code generation.
Added toolbar buttons to indicate status and provide quick access to the settings.
Refactoring for better performance
Added tokenizer and on KoboldCpp using its own tokenizer so it only adds to the prompt code that won't "push out" the main part of the prompt.
v1.5.1
Fixed a small issue that didn't change the indicator color on a failed connection.
v1.6.0
Reworked the way suggestions are generated.
Fixed an issue that prevented the extension to reconnect to the backend.
Added support for the "OogaBooga Text Generation WebUI" API.
Fixed multiple generation priority issues.
v1.6.1
After experimenting with LLaMa 3 it turned out that overflowing context messes up the generation beyond imaginable. The context that's sent in the generation request has now been adjusted to fit the context length.
v1.6.2
Fixed a disgusting issue that was overriding Context Length.
Fixed an issue that broke the generation pipeline.
