debian-mirror-gitlab/doc/user/project/repository/code_suggestions.md

stage	group	info	type
ModelOps	AI Assisted	To determine the technical writer assigned to the Stage/Group associated with this page, see https://about.gitlab.com/handbook/product/ux/technical-writing/#assignments	index, reference

Code Suggestions (Beta) (FREE SAAS)

• Introduced in GitLab 15.9 as Beta for early access Ultimate customers.
• Enabled as opt-in with GitLab 15.11 as Beta.
• Moved from GitLab Ultimate to GitLab Premium in 16.0.
• Moved from GitLab Premium to GitLab Free in 16.0.

WARNING: This feature is in Beta. Code Suggestions use generative AI to suggest code while you're developing. Due to high demand, this feature will have unscheduled downtime and code suggestions in VS Code may be delayed. Code Suggestions may produce low-quality or incomplete suggestions. Beta users should read about the known limitations. We look forward to hearing your feedback.

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 available in Visual Studio Code when you have the GitLab Workflow extension installed.

Supported languages

Code Suggestions may produce low-quality or incomplete suggestions. The best results from Code Suggestions are expected for these languages:

• C/C++
• C#
• Go
• Java
• JavaScript
• Python
• PHP
• Ruby
• Rust
• Scala
• TypeScript

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.

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 ({settings}), and select Extension Settings.
7. Enable GitLab > AI Assisted Code Suggestions.

Start typing and receive suggestions for your GitLab projects.

See an end-to-end demo: Get started with GitLab Code Suggestions in VS Code.

Code Suggestions data usage

Code Suggestions is a generative artificial intelligence (AI) model hosted on GitLab.com.

Your 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.

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.

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 the Google 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.

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 by using a GitLab personal access token with the read_api and read_user scopes.

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.

Internet connectivity

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 offline environment.

Stability and performance

This feature is currently in Beta. While the Code Suggestions inference API operates completely within the GitLab.com 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.

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.
• Add protections to limit personal data, insecure code, and other unwanted behavior that the model may have learned from training 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.