AI features based on 3rd-party integrations

Instructions for setting up GitLab Duo features in the local development environment

Required: Install AI gateway

Why: All Duo features route LLM requests through the AI gateway.

How: Follow these instructions to install the AI gateway with GDK. We recommend this route for most users.

You can also install AI gateway by:

  1. Cloning the repository directly.
  2. Running the server locally.

We only recommend this for users who have a specific reason for not running the AI gateway through GDK.

Required: Setup Licenses in GitLab-Rails

Why: GitLab Duo is available to Premium and Ultimate customers only. You likely want an Ultimate license for your GDK. Ultimate gets you access to all GitLab Duo features.

How:

Follow the process to obtain an EE license for your local instance and upload the license.

To verify that the license is applied, go to Admin area > Subscription and check the subscription plan.

Set up and run GDK

Option A: in SaaS (GitLab.com) Mode

Why: Most Duo features are available on GitLab.com first, so running in SaaS mode will ensure that you can access most features.

How:

Run the Rake task to set up Duo features for a group: 

GITLAB_SIMULATE_SAAS=1 bundle exec 'rake gitlab:duo:setup[test-group-name]'

gdk restart

Replace test-group-name with the name of any top-level group. Duo will be configured for that group. If the group doesn’t exist, it creates a new one.

Make sure the script succeeds. It prints error messages with links on how to resolve any errors. You can re-run the script until it succeeds.

In SaaS mode, membership to a group with Duo features enabled is what enables many AI features. Make sure that your test user is a member of the group with Duo features enabled (test-group-name).

This Rake task creates Duo Enterprise add-on attached to that group.

In case you need Duo Pro add-on attached, please use: 

GITLAB_SIMULATE_SAAS=1 bundle exec 'rake gitlab:duo:setup[test-group-name,duo_pro]'

Duo Pro add-on serves smaller scope of features. Usage of add-on depends on what features you want to use.

Option B: in Self-managed Mode

Why: If you want to test something specific to self-managed, such as Custom Models.

How:

Run the Rake task to set up Duo features for the instance: 

GITLAB_SIMULATE_SAAS=0 bundle exec 'rake gitlab:duo:setup_instance'

gdk restart

This Rake task creates Duo Enterprise add-on attached to your instance.

In case you need Duo Pro add-on attached, please use: 

GITLAB_SIMULATE_SAAS=0 bundle exec 'rake gitlab:duo:setup_instance[duo_pro]'

Duo Pro add-on serves smaller scope of features. Usage of add-on depends on what features you want to use.

Why: Setting this environment variable will allow the local GitLab instance to issue tokens itself, without syncing with CustomersDot first. With this set, you can skip the CustomersDot setup.

How: The following should be set in the env.runit file in your GDK root: 

# <GDK-root>/env.runit

export CLOUD_CONNECTOR_SELF_SIGN_TOKENS=1

You need to restart GDK to apply the change.

If you use CLOUD_CONNECTOR_SELF_SIGN_TOKENS=1, th root/admin user must have a seat assigned to receive a “Code completion test was successful” notification from the health check on the http://localhost:3000/admin/code_suggestions page.

Our customers (production environment) do not need to do that to run a Code Suggestions health check.

Why: you’ve completed all of the setup steps, now it’s time to confirm that GitLab Duo is actually working.

How:

After the setup is complete, you can test clients in GitLab-Rails to see if it can correctly reach to AI gateway:

  1. Run gdk start.
  2. Login to Rails console with gdk rails console.

  3. Talk to a model: 

    # Talk to Anthropic model
Gitlab::Llm::Anthropic::Client.new(User.first, unit_primitive: 'duo_chat').complete(prompt: "\n\nHuman: Hi, How are you?\n\nAssistant:")

# Talk to Vertex AI model
Gitlab::Llm::VertexAi::Client.new(User.first, unit_primitive: 'documentation_search').text_embeddings(content: "How can I create an issue?")

# Test `/v1/chat/agent` endpoint
Gitlab::Llm::Chain::Requests::AiGateway.new(User.first).request({prompt: [{role: "user", content: "Hi, how are you?"}]})
note
See this doc for registering unit primitives in Cloud Connector.

Optional: Enable authentication and authorization in AI gateway

Why: The AI gateway has authentication and authorization flow to verify if clients have permission to access the features. Auth is enforced in any live environments hosted by GitLab infra team. You may want to test this flow in your local development environment.

note
In development environments (for example: GDK), this process is disabled by default.

To enable authorization checks, set AIGW_AUTH__BYPASS_EXTERNAL to false in the application setting file (<GDK-root>/gitlab-ai-gateway/.env) in AI gateway.

Option 1: Use your GitLab instance as a provider

Why: this is the simplest method of testing authentication and reflects our setup on GitLab.com.

How: Assuming that you are running the AI gateway with GDK, apply the following configuration to GDK: 

# <GDK-root>/env.runit

export GITLAB_SIMULATE_SAAS=1

Update the application settings file in AI gateway: 

# <GDK-root>/gitlab-ai-gateway/.env

AIGW_AUTH__BYPASS_EXTERNAL=false
AIGW_GITLAB_URL=<your-gdk-url>

and gdk restart.

Option 2: Use your customersDot instance as a provider

Why: CustomersDot setup is required when you want to test or update functionality related to cloud licensing or if you are running GDK in non-SaaS mode.

note
This setup is challenging. There is an issue for discussing how to make it easier to test the customersDot integration locally. Until that is addressed, this setup process is time consuming and should be avoided if possible.

If you need to get customersDot working for your local GitLab Rails instance for any reason, reach out to #s_fulfillment_engineering in Slack. For questions around the integration of CDot with other systems to deliver AI use cases, reach out to #g_cloud_connector. assistance.

Help

Tips for local development

  1. When responses are taking too long to appear in the user interface, consider restarting Sidekiq by running gdk restart rails-background-jobs. If that doesn’t work, try gdk kill and then gdk start.
  2. Alternatively, bypass Sidekiq entirely and run the service synchronously. This can help with debugging errors as GraphQL errors are now available in the network inspector instead of the Sidekiq logs. To do that, temporarily alter the perform_for method in Llm::CompletionWorker class by changing perform_async to perform_inline.

Feature development (Abstraction Layer)

Feature flags

Apply the following feature flags to any AI feature work:

  • A general flag (ai_duo_chat_switch) that applies to all GitLab Duo Chat features. It’s enabled by default.
  • A general flag (ai_global_switch) that applies to all other AI features. It’s enabled by default.
  • A flag specific to that feature. The feature flag name must be different than the licensed feature name.

See the feature flag tracker epic for the list of all feature flags and how to use them.

Push feature flags to AI gateway

You can push feature flags to AI gateway. This is helpful to gradually rollout user-facing changes even if the feature resides in AI gateway. See the following example: 

# Push a feature flag state to AI gateway.
Gitlab::AiGateway.push_feature_flag(:new_prompt_template, user)

Later, you can use the feature flag state in AI gateway in the following way: 

from ai_gateway.feature_flags import is_feature_enabled

# Check if the feature flag "new_prompt_template" is enabled.
if is_feature_enabled('new_prompt_template'):
  # Build a prompt from the new prompt template
else:
  # Build a prompt from the old prompt template

IMPORTANT: At the cleaning up step, remove the feature flag in AI gateway repository before removing the flag in GitLab-Rails repository. If you clean up the flag in GitLab-Rails repository at first, the feature flag in AI gateway will be disabled immediately as it’s the default state, hence you might encounter a surprising behavior.

IMPORTANT: Cleaning up the feature flag in AI gateway will immediately distribute the change to all GitLab instances, including GitLab.com, Self-managed GitLab, and Dedicated.

Technical details:

  • When push_feature_flag runs on an enabled feature flag, the name of the flag is cached in the current context, which is later attached to the x-gitlab-enabled-feature-flags HTTP header when GitLab-Sidekiq/Rails sends requests to AI gateway.

  • When frontend clients (for example, VS Code Extension or LSP) request a User JWT (UJWT) for direct AI gateway communication, GitLab returns:

    • Public headers (including x-gitlab-enabled-feature-flags).
    • The generated UJWT (1-hour expiration).

Frontend clients must regenerate UJWT upon expiration. Backend changes such as feature flag updates through ChatOps render the header values to become stale. These header values are refreshed at the next UJWT generation.

Similarly, we also have push_frontend_feature_flag to push feature flags to frontend.

GraphQL API

To connect to the AI provider API using the Abstraction Layer, use an extendable GraphQL API called aiAction. The input accepts key/value pairs, where the key is the action that needs to be performed. We only allow one AI action per mutation request.

Example of a mutation: 

mutation {
  aiAction(input: {summarizeComments: {resourceId: "gid://gitlab/Issue/52"}}) {
    clientMutationId
  }
}

As an example, assume we want to build an “explain code” action. To do this, we extend the input with a new key, explainCode. The mutation would look like this: 

mutation {
  aiAction(
    input: {
      explainCode: { resourceId: "gid://gitlab/MergeRequest/52", code: "foo() { console.log() }" }
    }
  ) {
    clientMutationId
  }
}

The GraphQL API then uses the Anthropic Client to send the response.

How to receive a response

The API requests to AI providers are handled in a background job. We therefore do not keep the request alive and the Frontend needs to match the request to the response from the subscription.

caution
Determining the right response to a request can cause problems when only userId and resourceId are used. For example, when two AI features use the same userId and resourceId both subscriptions will receive the response from each other. To prevent this interference, we introduced the clientSubscriptionId.

To match a response on the aiCompletionResponse subscription, you can provide a clientSubscriptionId to the aiAction mutation.

  • The clientSubscriptionId should be unique per feature and within a page to not interfere with other AI features. We recommend to use a UUID.
  • Only when the clientSubscriptionId is provided as part of the aiAction mutation, it will be used for broadcasting the aiCompletionResponse.
  • If the clientSubscriptionId is not provided, only userId and resourceId are used for the aiCompletionResponse.

As an example mutation for summarizing comments, we provide a randomId as part of the mutation: 

mutation {
  aiAction(
    input: {
      summarizeComments: { resourceId: "gid://gitlab/Issue/52" }
      clientSubscriptionId: "randomId"
    }
  ) {
    clientMutationId
  }
}

In our component, we then listen on the aiCompletionResponse using the userId, resourceId and clientSubscriptionId ("randomId"): 

subscription aiCompletionResponse(
  $userId: UserID
  $resourceId: AiModelID
  $clientSubscriptionId: String
) {
  aiCompletionResponse(
    userId: $userId
    resourceId: $resourceId
    clientSubscriptionId: $clientSubscriptionId
  ) {
    content
    errors
  }
}

The subscription for Chat behaves differently.

To not have many concurrent subscriptions, you should also only subscribe to the subscription once the mutation is sent by using skip().

Current abstraction layer flow

The following graph uses VertexAI as an example. You can use different providers.

flowchart TD A[GitLab frontend] -->B[AiAction GraphQL mutation] B --> C[Llm::ExecuteMethodService] C --> D[One of services, for example: Llm::GenerateSummaryService] D -->|scheduled| E[AI worker:Llm::CompletionWorker] E -->F[::Gitlab::Llm::Completions::Factory] F -->G[`::Gitlab::Llm::VertexAi::Completions::...` class using `::Gitlab::Llm::Templates::...` class] G -->|calling| H[Gitlab::Llm::VertexAi::Client] H --> |response| I[::Gitlab::Llm::GraphqlSubscriptionResponseService] I --> J[GraphqlTriggers.ai_completion_response] J --> K[::GitlabSchema.subscriptions.trigger]

How to implement a new action

Implementing a new AI action will require changes across different components. We’ll use the example of wanting to implement an action that allows users to rewrite issue descriptions according to a given prompt.

1. Add your action to the Cloud Connector feature list

The Cloud Connector configuration stores the permissions needed to access your service, as well as additional metadata. If there’s still not an entry for your feature, you’ll need to add one in two places:

  • In the GitLab monolith: 
# ee/config/cloud_connector/access_data.yml

services:
  # ...
  rewrite_description:
    backend: 'gitlab-ai-gateway'
    bundled_with:
      duo_enterprise:
        unit_primitives:
          - rewrite_issue_description

# config/cloud_connector.yml

services:
  # ...
  rewrite_description:
    backend: 'gitlab-ai-gateway'
    bundled_with:
      duo_enterprise:
        unit_primitives:
          - rewrite_issue_description

For more information, see Cloud Connector: Configuration.

2. Create a prompt definition in the AI gateway

In the AI gateway project, create a new prompt definition under ai_gateway/prompts/definitions. Create a new subfolder corresponding to the name of your AI action, and a new YAML file for your prompt. Specify the model and provider you wish to use, and the prompts that will be fed to the model. You can specify inputs to be plugged into the prompt by using {}. 

# ai_gateway/prompts/definitions/rewrite_description/base.yml

name: Description rewriter
model:
  name: claude-3-sonnet-20240229
  params:
    model_class_provider: anthropic
prompt_template:
  system: |
    You are a helpful assistant that rewrites the description of resources. You'll be given the current description, and a prompt on how you should rewrite it. Reply only with your rewritten description.

    <description>{description}</description>

    <prompt>{prompt}</prompt>

If your AI action is part of a broader feature, the definitions can be organized in a tree structure: 

# ai_gateway/prompts/definitions/code_suggestions/generations/base.yml

name: Code generations
model:
  name: claude-3-sonnet-20240229
  params:
    model_class_provider: anthropic
...

To specify prompts for multiple models, use the name of the model as the filename for the definition: 

# ai_gateway/prompts/definitions/code_suggestions/generations/mistral.yml

name: Code generations
model:
  name: mistral
  params:
    model_class_provider: litellm
...

3. Create a Completion class

  1. Create a new completion under ee/lib/gitlab/llm/ai_gateway/completions/ and inherit it from the Base AI gateway Completion. 
# ee/lib/gitlab/llm/ai_gateway/completions/rewrite_description.rb

module Gitlab
  module Llm
    module AiGateway
      module Completions
        class RewriteDescription < Base
          def inputs
            { description: resource.description, prompt: prompt_message.content }
          end
        end
      end
    end
  end
end

4. Create a Service

  1. Create a new service under ee/app/services/llm/ and inherit it from the BaseService.
  2. The resource is the object we want to act on. It can be any object that includes the Ai::Model concern. For example it could be a Project, MergeRequest, or Issue. 
# ee/app/services/llm/rewrite_description_service.rb

module Llm
  class RewriteDescriptionService < BaseService
    extend ::Gitlab::Utils::Override

    override :valid
    def valid?
      super &&
        # You can restrict which type of resources your service applies to
        resource.to_ability_name == "issue" &&
        # Always check that the user is allowed to perform this action on the resource
        Ability.allowed?(user, :rewrite_description, resource)
    end

    private

    def perform
      schedule_completion_worker
    end
  end
end

5. Register the feature in the catalogue

Go to Gitlab::Llm::Utils::AiFeaturesCatalogue and add a new entry for your AI action. 

class AiFeaturesCatalogue
  LIST = {
    # ...
    rewrite_description: {
      service_class: ::Gitlab::Llm::AiGateway::Completions::RewriteDescription,
      feature_category: :ai_abstraction_layer,
      execute_method: ::Llm::RewriteDescriptionService,
      maturity: :experimental,
      self_managed: false,
      internal: false
    }
  }.freeze

Reuse the existing AI components for multiple models

We thrive optimizing AI components, such as prompt, input/output parser, tools/function-calling, for each LLM, however, diverging the components for each model could increase the maintenance overhead. Hence, it’s generally advised to reuse the existing components for multiple models as long as it doesn’t degrade a feature quality. Here are the rules of thumbs:

  1. Iterate on the existing prompt template for multiple models. Do NOT introduce a new one unless it causes a quality degredation for a particular model.
  2. Iterate on the existing input/output parsers and tools/functions-calling for multiple models. Do NOT introduce a new one unless it causes a quality degredation for a particular model.
  3. If a quality degredation is detected for a particular model, the shared component should be diverged for the particular model.

An example of this case is that we can apply Claude specific CoT optimization to the other models such as Mixtral as long as it doesn’t cause a quality degredation.

How to migrate an existing action to the AI gateway

AI actions were initially implemented inside the GitLab monolith. As part of our AI gateway as the Sole Access Point for Monolith to Access Models Epic we’re migrating prompts, model selection and model parameters into the AI gateway. This will increase the speed at which we can deliver improvements to self-managed users, by decoupling prompt and model changes from monolith releases. To migrate an existing action:

  1. Follow steps 1 through 3 on How to implement a new action.
  2. Modify the entry for your AI action in the catalogue to list the new completion class as the aigw_service_class. 
class AiFeaturesCatalogue
  LIST = {
    # ...
    generate_description: {
      service_class: ::Gitlab::Llm::Anthropic::Completions::GenerateDescription,
      aigw_service_class: ::Gitlab::Llm::AiGateway::Completions::GenerateDescription,
      prompt_class: ::Gitlab::Llm::Templates::GenerateDescription,
      feature_category: :ai_abstraction_layer,
      execute_method: ::Llm::GenerateDescriptionService,
      maturity: :experimental,
      self_managed: false,
      internal: false
    },
    # ...
  }.freeze
  1. Create prompt_migration_#{feature_name} feature flag (e.g prompt_migration_generate_description)

When the feature flag is enabled, the aigw_service_class will be used to process the AI action. Once you’ve validated the correct functioning of your action, you can remove the aigw_service_class key and replace the service_class with the new AiGateway::Completions class to make it the permanent provider.

For a complete example of the changes needed to migrate an AI action, see the following MRs:

Authorization in GitLab-Rails

We recommend to use policies to deal with authorization for a feature. Currently we need to make sure to cover the following checks:

Some basic authorization is included in the Abstraction Layer classes that are base classes for more specialized classes.

What needs to be included in the code:

  1. Check for feature flag compatibility: Gitlab::Llm::Utils::FlagChecker.flag_enabled_for_feature?(ai_action) - included in the Llm::BaseService class.
  2. Check if resource is authorized: Gitlab::Llm::Utils::Authorizer.resource(resource: resource, user: user).allowed? - also included in the Llm::BaseService class.
  3. Both of those checks are included in the ::Gitlab::Llm::FeatureAuthorizer.new(container: subject_container, feature_name: action_name).allowed?
  4. Access to AI features depend on several factors, such as: their maturity, if they are enabled on self-managed, if they are bundled within an add-on etc.
    • Example of policy not connected to the particular resource.
    • Example of policy connected to the particular resource.
note
For more information, see the GitLab AI gateway documentation about authentication and authorization in AI gateway.

Pairing requests with responses

Because multiple users’ requests can be processed in parallel, when receiving responses, it can be difficult to pair a response with its original request. The requestId field can be used for this purpose, because both the request and response are assured to have the same requestId UUID.

Caching

AI requests and responses can be cached. Cached conversation is being used to display user interaction with AI features. In the current implementation, this cache is not used to skip consecutive calls to the AI service when a user repeats their requests. 

query {
  aiMessages {
    nodes {
      id
      requestId
      content
      role
      errors
      timestamp
    }
  }
}

This cache is used for chat functionality. For other services, caching is disabled. You can enable this for a service by using the cache_response: true option.

Caching has following limitations:

  • Messages are stored in Redis stream.
  • There is a single stream of messages per user. This means that all services currently share the same cache. If needed, this could be extended to multiple streams per user (after checking with the infrastructure team that Redis can handle the estimated amount of messages).
  • Only the last 50 messages (requests + responses) are kept.
  • Expiration time of the stream is 3 days since adding last message.
  • User can access only their own messages. There is no authorization on the caching level, and any authorization (if accessed by not current user) is expected on the service layer.

Check if feature is allowed for this resource based on namespace settings

There is one setting allowed on root namespace level that restrict the use of AI features:

  • experiment_features_enabled

To check if that feature is allowed for a given namespace, call: 

Gitlab::Llm::StageCheck.available?(namespace, :name_of_the_feature)

Add the name of the feature to the Gitlab::Llm::StageCheck class. There are arrays there that differentiate between experimental and beta features.

This way we are ready for the following different cases:

  • If the feature is not in any array, the check will return true. For example, the feature is generally available.

To move the feature from the experimental phase to the beta phase, move the name of the feature from the EXPERIMENTAL_FEATURES array to the BETA_FEATURES array.

Implement calls to AI APIs and the prompts

The CompletionWorker will call the Completions::Factory which will initialize the Service and execute the actual call to the API. In our example, we will use VertexAI and implement two new classes: 

# /ee/lib/gitlab/llm/vertex_ai/completions/rewrite_description.rb

module Gitlab
  module Llm
    module VertexAi
      module Completions
        class AmazingNewAiFeature < Gitlab::Llm::Completions::Base
          def execute
            prompt = ai_prompt_class.new(options[:user_input]).to_prompt

            response = Gitlab::Llm::VertexAi::Client.new(user, unit_primitive: 'amazing_feature').text(content: prompt)

            response_modifier = ::Gitlab::Llm::VertexAi::ResponseModifiers::Predictions.new(response)

            ::Gitlab::Llm::GraphqlSubscriptionResponseService.new(
              user, nil, response_modifier, options: response_options
            ).execute
          end
        end
      end
    end
  end
end

# /ee/lib/gitlab/llm/vertex_ai/templates/rewrite_description.rb

module Gitlab
  module Llm
    module VertexAi
      module Templates
        class AmazingNewAiFeature
          def initialize(user_input)
            @user_input = user_input
          end

          def to_prompt
            <<~PROMPT
            You are an assistant that writes code for the following context:

            context: #{user_input}
            PROMPT
          end
        end
      end
    end
  end
end

Because we support multiple AI providers, you may also use those providers for the same example: 

Gitlab::Llm::VertexAi::Client.new(user, unit_primitive: 'your_feature')
Gitlab::Llm::Anthropic::Client.new(user, unit_primitive: 'your_feature')

Monitoring

Logs

Overview

In addition to standard logging in the GitLab Rails Monolith instance, specialized logging is available for features based on large language models (LLMs).

Logged events

Currently logged events are documented here.

Implementation

Logger Class

To implement LLM-specific logging, use the Gitlab::Llm::Logger class.

Privacy Considerations

Important: User inputs and complete prompts containing user data must not be logged unless explicitly permitted.

Feature Flag

A feature flag named expanded_ai_logging controls the logging of sensitive data. Use the conditional_info helper method for conditional logging based on the feature flag status:

  • If the feature flag is enabled for the current user, it logs the information on info level (logs are accessible in Kibana).
  • If the feature flag is disabled for the current user, it logs the information on info level, but without optional parameters (logs are accessible in Kibana, but only obligatory fields).

Best Practices

When implementing logging for LLM features, consider the following:

  • Identify critical information for debugging purposes.
  • Ensure compliance with privacy requirements by not logging sensitive user data without proper authorization.
  • Use the conditional_info helper method to respect the expanded_ai_logging feature flag.
  • Structure your logs to provide meaningful insights for troubleshooting and analysis.

Example Usage 

# including concern that handles logging
include Gitlab::Llm::Concerns::Logger

# Logging potentially sensitive information
log_conditional_info(user, message:"User prompt processed", event_name: 'ai_event', ai_component: 'abstraction_layer', prompt: sanitized_prompt)

# Logging application error information
log_error(user, message: "System application error", event_name: 'ai_event', ai_component: 'abstraction_layer', error_message: sanitized_error_message)

Important: Please familiarize yourself with our Data Retention Policy and remember to make sure we are not logging user input and LLM-generated output.

Security

Refer to the secure coding guidelines for Artificial Intelligence (AI) features.

Model Migration Process

Introduction

LLM models are constantly evolving, and GitLab needs to regularly update our AI features to support newer models. This guide provides a structured approach for migrating AI features to new models while maintaining stability and reliability.

Purpose

Provide a comprehensive guide for migrating AI models within GitLab.

Expected Duration

Model migrations typically follow these general timelines:

  • Simple Model Updates (Same Provider): 2-3 weeks
    • Example: Upgrading from Claude Sonnet 3.5 to 3.6
    • Involves model validation, testing, and staged rollout
    • Primary focus on maintaining stability and performance
    • Can sometimes be expedited when urgent, but 2 weeks is standard
  • Complex Migrations: 1-2 months (full milestone or longer)
    • Example: Adding support for a new provider like AWS Bedrock
    • Example: Major version upgrades with breaking changes (e.g., Claude 2 to 3)
    • Requires significant API integration work
    • May need infrastructure changes
    • Extensive testing and validation required

Timeline Factors

Several factors can impact migration timelines:

  • Current system stability and recent incidents
  • Resource availability and competing priorities
  • Complexity of behavioral changes in new model
  • Scale of testing required
  • Feature flag rollout strategy

Best Practices

  • Always err on the side of caution with initial timeline estimates
  • Use feature flags for gradual rollouts to minimize risk
  • Plan for buffer time to handle unexpected issues
  • Communicate conservative timelines externally while working to deliver faster
  • Prioritize system stability over speed of deployment
note
While some migrations can technically be completed quickly, we typically plan for longer timelines to ensure proper testing and staged rollouts. This approach helps maintain system stability and reliability.

Scope

Applicable to all AI model-related teams at GitLab. We currently only support using Anthropic and Google Vertex models, with plans to support AWS Bedrock models in the future.

Prerequisites

Before starting a model migration:

  • Create an issue under the AI Model Version Migration Initiative epic with the following:
    • Label with group::ai framework
    • Document any known behavioral changes or improvements in the new model
    • Include any breaking changes or compatibility issues
    • Reference any model provider documentation about the changes

  • Verify the new model is supported in our current AI-Gateway API specification by:

    • Check model definitions in AI gateway:
      • For LiteLLM models: ai_gateway/models/v2/container.py
      • For Anthropic models: ai_gateway/models/anthropic.py
      • For new providers: Create a new model definition file in ai_gateway/models/
    • Verify model configurations:
      • Model enum definitions
      • Stop tokens
      • Timeout settings
      • Completion type (text or chat)
      • Max token limits
    • Testing the model locally in AI gateway:
    • If the model isn’t supported, create an issue in the AI gateway repository to add support
    • Review the provider’s API documentation for any breaking changes:
  • Ensure you have access to testing environments and monitoring tools
  • Complete model evaluation using the Prompt Library
note
Documentation of model changes is crucial for tracking the impact of migrations and helping with future troubleshooting. Always create an issue to track these changes before beginning the migration process.

Migration Tasks

Migration Tasks for Anthropic Model

  • Optional - Investigate if the new model is supported within our current AI-Gateway API specification. This step can usually be skipped. However, sometimes to support a newer model, we may need to accommodate a new API format.
  • Add the new model to our available models list.
  • Change the default model in our AI-Gateway client. Please place the change around a feature flag. We may need to quickly rollback the change.
  • Update the model definitions in AI gateway following the prompt definition guidelines Note: While we’re moving toward AI gateway holding the prompts, feature flag implementation still requires a GitLab release.

Migration Tasks for Vertex Models

Work in Progress

Feature Flag Process

Implementation Steps

For implementing feature flags, refer to our Feature Flags Development Guidelines.

note
Feature flag implementations will affect self-hosted cloud-connected customers. These customers won’t receive the model upgrade until the feature flag is removed from the AI gateway codebase, as they won’t have access to the new GitLab release.

Model Selection Implementation

The model selection logic should be implemented in:

  • AI gateway client (ee/lib/gitlab/llm/chain/requests/ai_gateway.rb)
  • Model definitions in AI gateway
  • Any custom implementations in specific features that override the default model

Rollout Strategy

  • Enable the feature flag for a small percentage of users/groups initially
  • Monitor performance metrics and error rates using:
  • Gradually increase the rollout percentage
  • If issues arise, quickly disable the feature flag to rollback to the previous model
  • Once stability is confirmed, remove the feature flag and make the migration permanent

For more details on monitoring during migrations, see the Monitoring and Metrics section below.

Scope of Work

AI Features to Migrate

  • Duo Chat Tools:
    • ci_editor_assistant/prompts/anthropic.rb - CI Editor
    • gitlab_documentation/executor.rb - GitLab Documentation
    • epic_reader/prompts/anthropic.rb - Epic Reader
    • issue_reader/prompts/anthropic.rb - Issue Reader
    • merge_request_reader/prompts/anthropic.rb - Merge Request Reader
  • Chat Slash Commands:
    • refactor_code/prompts/anthropic.rb - Refactor
    • write_tests/prompts/anthropic.rb - Write Tests
    • explain_code/prompts/anthropic.rb - Explain Code
    • explain_vulnerability/executor.rb - Explain Vulnerability
  • Experimental Tools:
    • Summarize Comments Chat
    • Fill MR Description

Testing and Validation

Model Evaluation

The ai-model-validation team created the following library to evaluate the performance of prompt changes as well as model changes. The Prompt Library README.MD provides details on how to evaluate the performance of AI features.

Another use-case for running chat evaluation is during feature development cycle. The purpose is to verify how the changes to the code base and prompts affect the quality of chat responses before the code reaches the production environment.

For evaluation in merge request pipelines, we use:

Seed project and group resources for testing and evaluation

To seed project and group resources for testing and evaluation, run the following command: 

SEED_GITLAB_DUO=1 FILTER=gitlab_duo bundle exec rake db:seed_fu

This command executes the development seed file for GitLab Duo, which creates gitlab-duo group in your GDK.

This command is responsible for seeding group and project resources for testing GitLab Duo features. It’s mainly used by the following scenarios:

  • Developers or UX designers have a local GDK but don’t know how to set up the group and project resources to test a feature in UI.
  • Evaluators (e.g. CEF) have input dataset that refers to a group or project resource e.g. (Summarize issue #123 requires a corresponding issue record in PosstgreSQL)

Currently, the input dataset of evaluators and this development seed file are managed seaprately. To ensure that the integration keeps working, this seeder has to create the same group/project resources every time. For example, ID and IID of the inserted PostgreSQL records must be the same every time we run this seeding process.

These fixtures are depended by the following projects:

See this architecture doc for more information.

Local Development

A valuable tool for local development to ensure the changes are correct outside of unit tests is to use LangSmith for tracing. The tool allows you to trace LLM calls within Duo Chat to verify the LLM tool is using the correct model.

To prevent regressions, we also have CI jobs to make sure our tools are working correctly. For more details, see the Duo Chat testing section.

Monitoring and Metrics

Monitor the following during migration: