• Retrieve GitLab Documentation
  • Setup in Vertex AI Search
  • AI Gateway API
  • Proof of Concept
    • Evaluation score
  • Estimated Time of Completion

Vertex AI Search

Retrieve GitLab Documentation

• Statistics (as of January 2024):
  • Date type: Markdown (Unstructured) written in natural language
  • Date access level: Green (No authorization required)
  • Data source: https://gitlab.com/gitlab-org/gitlab/-/blob/master/doc
  • Data size: approx. 56,000,000 bytes. 2194 pages.
  • Service: https://docs.gitlab.com/ (source repo
  • Example of user input: “How do I create an issue?”
  • Example of expected AI-generated response: “To create an issue:\n\nOn the left sidebar, select Search or go to and find your project.\n\nOn the left sidebar, select Plan > Issues, and then, in the upper-right corner, select New issue.”

The GitLab documentation is the SSoT service to serve GitLab documentation for SaaS (both GitLab.com and Dedicated) and Self-managed. When a user accesses to a documentation link in GitLab instance, they are redirected to the service since 16.0 (except air-gapped solutions). In addition, the current search backend of docs.gitlab.com needs to transition to Vertex AI Search. See this issue (GitLab member only) for more information.

We introduce a new semantic search API powered by Vertex AI Search for the documentation tool of GitLab Duo Chat.

Setup in Vertex AI Search

We create a search app for each GitLab versions. These processes will likely be automated in the GitLab Documentation project by CI/CD pipelines.

1. Create a new Bigquery table e.g. gitlab-docs-latest or gitlab-docs-v16.4
2. Download documents from repositories (e.g. gitlab-org/gitlab/doc, gitlab-org/gitlab-runner/docs, gitlab-org/omnibus-gitlab/doc).
3. Split them by Markdown headers and generate metadata (e.g. URL and title).
4. Insert rows into the Bigquery table.
5. Create a search app

See this notebook for more implementation details. The data of the latest version will be refreshed by a nightly build with Data Store API.

AI Gateway API

API design is following the existing patterns in AI Gateway.

POST /v1/search/docs

{
  "type": "search",
  "metadata": {
    "source": "GitLab EE",
    "version": "16.3"         // Used for switching search apps for older GitLab instances
  },
  "payload": {
    "query": "How can I create an issue?",
    "params": {               // Params for Vertex AI Search
      "page_size": 10,
      "filter": "",
    },
    "provider": "vertex-ai"
  }
}

The response will include the search results. For example:

{
  "response": {
    "results": [
      {
        "id": "d0454e6098773a4a4ebb613946aadd89",
        "content": "\nTo create an issue from a group:  \n1. On the left sidebar, ...",
        "metadata": {
          "Header1": "Create an issue",
          "Header2": "From a group",
          "url": "https://docs.gitlab.com/ee/user/project/issues/create_issues.html"
        }
      }
    ]
  },
  "metadata": {
    "provider": "vertex-ai"
  }
}

See SearchRequest and SearchResponse for Vertex AI API specs.

Proof of Concept

• GitLab-Rails MR
• AI Gateway MR
• Vertex AI Search service
• Google Colab notebook
• Demo video (Note: In this video, Website URLs are used as data source).

Evaluation score

Here is the evaluation scores generated by Prompt Library.

Estimated Time of Completion

• Milestone N:
  • Setup in Vertex AI Search with CI/CD automation.
  • Introduce /v1/search/docs endpoint in AI Gateway.
  • Updates the retrieval logic in GitLab-Rails.
  • Feature flag clean up.

Total milestones: 1