In March, we introduced fundamental improvements in the Refact.ai server (self-hosting and Enterprise): now you can train several fine-tunes and serve it to different teams inside your company, without any loss in speed! Fine-tuning now supports StarCoder2 models. We added chat system prompt and toolbox customization. Continue reading for more details!
New plan: Refact.ai Teams
Refact.ai Teams, a simplified version of Enterprise plan, is hosted on our cloud and provides per-user statistics and multi-user access control. It’s perfect for companies wanting to test-drive Refact.ai Enterprise without a full server deployment. Curious? Read our blog post about Teams and give it a try!
More commands in AI Toolbox (+custom)
Press F1 in VS code to access the expanded AI Toolbox:
New Commands: /help, /gen, /edit
Custom Commands: create your own prompt to solve mundane tasks faster. Check the how-to in our docs! Enterprise users can also assign team-specific commands.
/edit command in AI Toolbox
Custom System Prompt in Chat
Set up your own system prompt as the default for chat in both VS Code and JetBrains settings.
Example of system prompt you can create in a YAML file
With Refact.ai Enterprise, you can also allocate different default prompts to various teams.
New Models
Added in March and available for fine-tuning:
Starcoder2 3B, 7B, 15B
Deepseek-coder vLLM variants.
Enterprise and Self-hosted Updates
Docker v1.5.0:
It’s easier to train and compare fine-tune runs:
Set up multiple projects with different source files
Run several multi-GPU fine-tune runs simultaneously
The Enterprise version now has team preferences
Set fine-tune for chat and code completion models per team
Set custom default chat system prompt per team
Refact.ai model training
Plugins Updates
Force completion shortcut now increases the temperature from the second invocation, to generate more diverse completions (default shortcuts are: ⌥ + Space in VS Code; ⌥ + / in JetBrains)
You can choose chat system prompt when starting a new chat (VS Code, JetBrains coming up)
To get maximum hardware acceleration for per-team fine-tunes, we had to change the format a finetune is stored on disk and used. This requires to re-train the finetune you might already have.
Don’t forget to update your plugins! Old ones will still work in a backward-compatibility mode, but all recent features only work with the latest plugins.
We hope you liked the March updates of Refact.ai AI coding assistant! Our team is making it the best copilot alternative with fine-tuning and features, boosting developer productivity.