• Group level setting
• Enable Code Suggestions in VS Code
• Code Suggestions Data Usage
  • Overview
    • Progressive enhancement
    • Off by default
    • Generating suggestions
  • Stability and performance
• Data privacy
  • Model accuracy and quality
  • Training data
• Known limitations
• Feedback

Code Suggestions (Beta)

Use Code Suggestions to write code more efficiently by viewing code suggestions as you type. Depending on the cursor position, the extension either:

• Provides entire code snippets, like generating functions.
• Completes the current line.

To accept a suggestion, press Tab.

Code Suggestions are supported in Visual Studio Code with the GitLab Workflow extension. GitLab plans to support the new GitLab WebIDE in an upcoming release in the future.

Code Suggestions may produce low-quality or incomplete suggestions. Beta users should read about the known limitations. The best results from Code Suggestions are expected for these six languages:

• C
• C++
• Go
• Java
• JavaScript
• Python

Suggestions may be mixed for other languages. Using natural language code comments to request completions may also not function as expected.

GitLab is continuously improving the model and expects to support an additional seven languages soon, as well as natural language code comments.

Usage of Code Suggestions is governed by the GitLab Testing Agreement. Learn about data usage when using Code Suggestions.

Group level setting

Group owners can enable Code Suggestions for all projects in a group by using the group level Code Suggestions setting.

Enable Code Suggestions in VS Code

Prerequisites:

• Your group owner must enable the group level code suggestions setting.
• You must have a personal access token with the read_api and read_user scopes.

To enable Code Suggestions in VS Code:

1. Download and configure the GitLab Workflow extension for Visual Studio Code.
2. In GitLab: Add Account to VS Code on Mac, add your GitLab work account to the VS Code extension:
   • In macOS, press Shift + Command + P.
   • In Windows, press Shift + Control + P.
3. Provide your GitLab instance URL. A default is provided.
4. Provide your personal access token.
5. After your GitLab account connects successfully, in the left sidebar, select Extensions.
6. Find the GitLab workflow extension, select Settings (), and select Extension Settings.
7. Enable GitLab > AI Assisted Code Suggestions.

Start typing and receive suggestions for your GitLab projects.

Code Suggestions Data Usage

Overview

Code Suggestions is a generative artificial intelligence (AI) model hosted on GitLab.com that can empower your developers to code more efficiently by suggesting code as they type.

The personal access token enables a secure API connection to GitLab.com. This API connection securely transmits a context window from VS Code to the Code Suggestions ML model for inference, and the generated suggestion is transmitted back to VS Code.

Progressive enhancement

This feature is designed as a progressive enhancement to the existing VS Code GitLab Workflow plugin. Code Suggestions offer a completion if the machine learning engine can generate a recommendation. In the event of a connection issue or model inference failure, the feature gracefully degrades. Code Suggestions do not prevent you from writing code in VS Code.

Off by default

Code Suggestions are off by default and require a group owner to enable the feature with a group-level setting.

After the group level setting is enabled, Developers using Visual Studio Code with the GitLab Workflow extension can connect to GitLab.com via a GitLab personal access token with the read_api and read_user scopes.

Generating suggestions

Once configured by a developer in VS Code. The personal access token enables a secure API connection to GitLab.com. This API connection securely transmits a context window from VS Code to the Code Suggestions ML model for inference, and the generated suggestion is transmitted back to VS Code.

Code Suggestions only work when you have internet connectivity and can access GitLab.com. Code Suggestions are not available for self-managed customers, nor customers operating within an air-gapped environment.

Stability and performance

This feature is currently in Beta. While the Code Suggestions inference API operates completely within GitLab.com’s enterprise infrastructure, we expect a high demand for this Beta feature, which may cause degraded performance or unexpected downtime of the feature. We have built this feature to gracefully degrade and have controls in place to allow us to mitigate abuse or misuse. GitLab may disable this feature for any or all customers at any time at our discretion.

Data privacy

Code Suggestions operate completely in the GitLab.com infrastructure, providing the same level of security as any other feature of GitLab.com, and processing any personal data in accordance with our Privacy Statement.

No new additional data is collected to enable this feature. The content of your GitLab hosted source code is not used as training data. Source code inference against the Code Suggestions model is not used to re-train the model. Your data also never leaves GitLab.com. All training and inference is done in GitLab.com infrastructure.

Read more about the security of GitLab.com.

Model accuracy and quality

While in Beta, Code Suggestions can generate low-quality, incomplete, and possibly insecure code. We strongly encourage all beta users to leverage GitLab native Code Quality Scanning and Security Scanning capabilities.

GitLab uses a customized open source dataset to fine-tune the model to support multiple languages. Based on the languages you code in, GitLab routes the request to a targeted inference and prompt engine to get relevant suggestions.

GitLab is actively refining these models to improve the quality of recommendations, add support for more languages, and add protections to limit personal data, insecure code, and other unwanted behavior that the model may have learned from training data.

Training data

Code Suggestions uses open source pre-trained base models from the CodeGen family including CodeGen-MULTI and CodeGen-NL. We then re-train and fine-tune these base models with a customized open source dataset to enable multi-language support and additional use cases. This customized dataset contains non-preprocessed open source code in 13 programming languages from The Pile and Google’s BigQuery source code dataset. We then process this raw dataset against heuristics that aim to increase the quality of the dataset.

The Code Suggestions model is not trained on GitLab customer data.

Known limitations

While in Beta, we are working on improving the accuracy of overall generated content. However, Code Suggestions may generate suggestions that are:

• Low-quality
• Incomplete
• Produce failed pipelines
• Insecure code
• Offensive or insensitive

We are also aware of specific situations that can produce unexpected or incoherent results including:

• Suggestions based on natural language code comments.
• Suggestions that mixed programming languages in unexpected ways.

Feedback

Report issues in the feedback issue.