Semantic code search

Tier: Premium, Ultimate
Add-on: GitLab Duo Core, Pro, or Enterprise
Offering: GitLab.com, GitLab Self-Managed, GitLab Dedicated
Status: Beta

Version history

Introduced as a beta in GitLab 18.7.

Semantic code search uses AI to find relevant code snippets in your repository based on meaning rather than keyword matching.

Semantic code search converts your codebase into vector embeddings and stores these embeddings in a vector database. Your search query is also converted into an embedding and then compared against your code embeddings to find the most semantically similar results. This approach finds relevant code even when keywords do not match.

Improvements to this feature are proposed in epic 18018 and epic 20110.

Prerequisites

Have one of these configured:
- Access to the GitLab AI Gateway.
- A self-hosted AI Gateway with access to the Vertex AI text-embedding-005 model for embedding generation.
Have these features turned on:
- For GitLab.com, experiment features for your top-level namespace.
- For GitLab Self-Managed, GitLab Duo experiment and beta features for the instance.
Have GitLab Duo turned on for your project.
Have a supported vector store configured:
- Elasticsearch 8.0 and later.
- OpenSearch 2.0 and later.

Enable semantic code search

With the UI

If your GitLab instance uses Elasticsearch or OpenSearch for advanced search, you can enable semantic code search by connecting to the same cluster:

In the upper-right corner, select Admin.
Select Settings > Search.
Expand Semantic search.
Select Connect to the advanced search cluster.

With the Rails console

To create a custom vector store connection for Elasticsearch or OpenSearch, in the Rails console, create a connection with adapter and options.

Elasticsearch

connection = Ai::ActiveContext::Connection.create!(
  name: "elasticsearch",
  options: { url: ["http://your-elasticsearch-url:9200"] },
  adapter_class: "ActiveContext::Databases::Elasticsearch::Adapter"
)
connection.activate!

Connection options:

Option	Type	Required	Default	Description
`url`	array of strings	Yes	None	Array of URLs for your Elasticsearch cluster (for example, `["http://localhost:9200"]`).
`client_adapter`	string	No	`typhoeus`	HTTP adapter to use. Possible values are `typhoeus` and `net_http`.
`client_request_timeout`	integer	No	`30`	Request timeout in seconds.
`retry_on_failure`	integer	No	`0`	Number of retries on failure.
`debug`	boolean	No	`false`	Enables debug logging.

OpenSearch

connection = Ai::ActiveContext::Connection.create!(
  name: "opensearch",
  options: { url: ["http://your-opensearch-url:9200"] },
  adapter_class: "ActiveContext::Databases::Opensearch::Adapter"
)
connection.activate!

Connection options:

Option	Type	Required	Default	Description
`url`	array of strings	Yes	None	Array of URLs for your OpenSearch cluster (for example, `["http://localhost:9200"]`).
`client_adapter`	string	No	`typhoeus`	HTTP adapter to use. Possible values are `typhoeus` and `net_http`.
`client_request_timeout`	integer	No	`30`	Request timeout in seconds.
`retry_on_failure`	integer	No	`0`	Number of retries on failure.
`debug`	boolean	No	`false`	Enables debug logging.
`aws`	boolean	No	`false`	Enables AWS Signature Version 4 signing.
`aws_region`	string	No	None	AWS region for your OpenSearch domain.
`aws_access_key`	string	No	None	AWS access key ID.
`aws_secret_access_key`	string	No	None	AWS secret access key.

Use semantic code search

Semantic code search is available as a GitLab MCP server tool. For more information about how to use this tool, see semantic_code_search.

When you first use semantic code search in a GitLab project:

Your repository code is indexed and converted into vector embeddings.
These embeddings are stored in your configured vector store.
Updates are processed incrementally when code is merged to the default branch.

Initial indexing might take a few minutes depending on your repository size.