25 KiB

Raw Blame History

stage	group	info
ModelOps	AI Framework	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

AI features based on 3rd-party integrations

Introduced in GitLab 15.11.

Features

Async execution of the long running API requests
- GraphQL Action starts the request
- Background workers execute
- GraphQL subscriptions deliver results back in real time
Abstraction for
- OpenAI
- Google Vertex AI
- Anthropic
Rate Limiting
Circuit Breaker
Multi-Level feature flags
License checks on group level
Snowplow execution tracking
Tracking of Token Spent on Prometheus
Configuration for Moderation check of inputs
Automatic Markdown Rendering of responses
Centralised Group Level settings for experiment and 3rd party
Experimental API endpoints for exploration of AI APIs by GitLab team members without the need for credentials
- OpenAI
- Google Vertex AI
- Anthropic

Feature flags

Apply the following two feature flags to any AI feature work:

A general that applies to all AI features.
A flag specific to that feature. The feature flag name must be different than the licensed feature name.

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

Implement a new AI action

To implement a new AI action, connect to the preferred AI provider. You can connect to this API using either the:

Experimental REST API.
Abstraction layer.

All AI features are experimental.

Test AI features locally

NOTE: Use this snippet for help automating the following section.

Enable the required general feature flags:

Feature.enable(:ai_related_settings)
Feature.enable(:openai_experimentation)
Feature.enable(:tofa_experimentation_main_flag)
Feature.enable(:anthropic_experimentation)

Simulate the GDK to simulate SaaS and ensure the group you want to test has an Ultimate license
Enable Experimental features and Third-party AI services
1. Go to the group with the Ultimate license
2. Group Settings > General -> Permissions and group features
3. Enable Experiment features
4. Enable Third-party AI services
Enable the specific feature flag for the feature you want to test
Set the required access token. To receive an access token:
1. For Vertex, follow the instructions below.
2. For all other providers, like Anthropic or OpenAI, create an access request where @m_gill, @wayne, and @timzallmann are the tech stack owners.

Set up the embedding database

NOTE: Use this snippet for help automating the following section.

For features that use the embedding database, additional setup is needed.

Enable pgvector in GDK

Enable the embedding database in GDK

  gdk config set gitlab.rails.databases.embedding.enabled true

Run gdk reconfigure
Run database migrations to create the embedding database

Set up GitLab Duo Chat

NOTE: Use this snippet for help automating the following section.

Enable Anthropic API features.
Enable OpenAI support.
Ensure the embedding database is configured.

Enable feature specific feature flag.

Feature.enable(:gitlab_duo)
Feature.enable(:tanuki_bot)
Feature.enable(:ai_redis_cache)

Ensure that your current branch is up-to-date with master.
To access the GitLab Duo Chat interface, in the lower-left corner of any page, select Help and Ask GitLab Duo Chat.

Tips for local development

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.
Alternatively, bypass Sidekiq entirely and run the chat service synchronously. This can help with debugging errors as GraphQL errors are now available in the network inspector instead of the Sidekiq logs.

diff --git a/ee/app/services/llm/chat_service.rb b/ee/app/services/llm/chat_service.rb
index 5fa7ae8a2bc1..5fe996ba0345 100644
--- a/ee/app/services/llm/chat_service.rb
+++ b/ee/app/services/llm/chat_service.rb
@@ -5,7 +5,7 @@ class ChatService < BaseService
     private

     def perform
-      worker_perform(user, resource, :chat, options)
+      worker_perform(user, resource, :chat, options.merge(sync: true))
     end

     def valid?

Working with GitLab Duo Chat

Prompts are the most vital part of GitLab Duo Chat system. Prompts are the instructions sent to the Large Language Model to perform certain tasks.

The state of the prompts is the result of weeks of iteration. If you want to change any prompt in the current tool, you must put it behind a feature flag.

If you have any new or updated prompts, ask members of AI Framework team to review, because they have significant experience with them.

Setup for GitLab documentation chat (legacy chat)

To populate the embedding database for GitLab chat:

Open a rails console
Run this script to populate the embedding database

Contributing to GitLab Duo Chat

The Chat feature uses a zero-shot agent that includes a system prompt explaining how the large language model should interpret the question and provide an answer. The system prompt defines available tools that can be used to gather information to answer the user's question.

The zero-shot agent receives the user's question and decides which tools to use to gather information to answer it. It then makes a request to the large language model, which decides if it can answer directly or if it needs to use one of the defined tools.

The tools each have their own prompt that provides instructions to the large language model on how to use that tool to gather information. The tools are designed to be self-sufficient and avoid multiple requests back and forth to the large language model.

After the tools have gathered the required information, it is returned to the zero-shot agent, which asks the large language model if enough information has been gathered to provide the final answer to the user's question.

Adding a new tool

To add a new tool:

Create files for the tool in the ee/lib/gitlab/llm/chain/tools/ folder. Use existing tools like issue_identifier or resource_reader as a template.
Write a class for the tool that includes:
- Name and description of what the tool does
- Example questions that would use this tool
- Instructions for the large language model on how to use the tool to gather information - so the main prompts that this tool is using.
Test and iterate on the prompt using RSpec tests that make real requests to the large language model.
- Prompts require trial and error, the non-deterministic nature of working with LLM can be surprising.
- Anthropic provides good guide on working on prompts.
Implement code in the tool to parse the response from the large language model and return it to the zero-shot agent.
Add the new tool name to the tools array in ee/lib/gitlab/llm/completions/chat.rb so the zero-shot agent knows about it.
Add tests by adding questions to the test-suite for which the new tool should respond to. Iterate on the prompts as needed.

The key things to keep in mind are properly instructing the large language model through prompts and tool descriptions, keeping tools self-sufficient, and returning responses to the zero-shot agent. With some trial and error on prompts, adding new tools can expand the capabilities of the chat feature.

There are available short videos covering this topic.

Debugging

To gather more insights about the full request, use the Gitlab::Llm::Logger file to debug logs. The default logging level on production is INFO and must not be used to log any data that could contain personal identifying information.

To follow the debugging messages related to the AI requests on the abstraction layer, you can use:

export LLM_DEBUG=1
gdk start
tail -f log/llm.log

Configure GCP Vertex access

In order to obtain a GCP service key for local development, please follow the steps below:

Create a sandbox GCP environment by visiting this page and following the instructions, or by requesting access to our existing group environment by using this template.
In the GCP console, go to IAM & Admin > Service Accounts and click on the "Create new service account" button
Name the service account something specific to what you're using it for. Select Create and Continue. Under Grant this service account access to project, select the role Vertex AI User. Select Continue then Done
Select your new service account and Manage keys > Add Key > Create new key. This will download the private JSON credentials for your service account.
Open the Rails console. Update the settings to:

Gitlab::CurrentSettings.update(vertex_ai_credentials: File.read('/YOUR_FILE.json'))

# Note: These credential examples will not work locally for all models
Gitlab::CurrentSettings.update(vertex_ai_host: "<root-domain>") # Example: us-central1-aiplatform.googleapis.com
Gitlab::CurrentSettings.update(vertex_ai_project: "<project-id>") # Example: cloud-large-language-models

Internal team members can use this snippet for help configuring these endpoints.

Configure OpenAI access

Gitlab::CurrentSettings.update(openai_api_key: "<open-ai-key>")

Configure Anthropic access

Feature.enable(:anthropic_experimentation)
Gitlab::CurrentSettings.update!(anthropic_api_key: <insert API key>)

Testing GitLab Duo Chat with predefined questions

Because success of answers to user questions in GitLab Duo Chat heavily depends on toolchain and prompts of each tool, it's common that even a minor change in a prompt or a tool impacts processing of some questions. To make sure that a change in the toolchain doesn't break existing functionality, you can use following commands to validate answers to some predefined questions:

Rake task which iterates through questions defined in CSV file and checks tools used for evaluating each question.

rake gitlab:llm:zero_shot:test:questions[<issue_url>]

RSpec which iterates through resource-specific questions on predefined resources:

ANTHROPIC_API_KEY='<key>' REAL_AI_REQUEST=1 rspec ee/spec/lib/gitlab/llm/chain/agents/zero_shot/executor_spec.rb

Experimental REST API

Use the experimental REST API endpoints to quickly experiment and prototype AI features.

The endpoints are:

https://gitlab.example.com/api/v4/ai/experimentation/openai/completions
https://gitlab.example.com/api/v4/ai/experimentation/openai/embeddings
https://gitlab.example.com/api/v4/ai/experimentation/openai/chat/completions
https://gitlab.example.com/api/v4/ai/experimentation/anthropic/complete
https://gitlab.example.com/api/v4/ai/experimentation/tofa/chat

These endpoints are only for prototyping, not for rolling features out to customers. The experimental endpoint is only available to GitLab team members on production. Use the GitLab API token to authenticate.

Abstraction layer

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 OpenAI Client to send the response.

Remember that other clients are available and you should not use OpenAI.

How to receive a response

As the OpenAI API requests are handled in a background job, we do not keep the request alive and the response is sent through the aiCompletionResponse subscription:

subscription aiCompletionResponse($userId: UserID, $resourceId: AiModelID!) {
  aiCompletionResponse(userId: $userId, resourceId: $resourceId) {
    responseBody
    errors
  }
}

WARNING: You should only subscribe to the subscription once the mutation is sent. If multiple subscriptions are active on the same page, they currently all receive updates as our identifier is the user and the resource. To mitigate this, you should only subscribe when the mutation is sent. You can use [skip()](You can use skip()) for this case. To prevent this problem in the future, we implement a request identifier.

Current abstraction layer flow

The following graph uses OpenAI 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::OpenAi::Completions::...` class using `::Gitlab::Llm::OpenAi::Templates::...` class]
G -->|calling| H[Gitlab::Llm::OpenAi::Client]
H --> |response| I[::Gitlab::Llm::OpenAi::ResponseService]
I --> J[GraphqlTriggers.ai_completion_response]
J --> K[::GitlabSchema.subscriptions.trigger]

CircuitBreaker

The CircuitBreaker concern is a reusable module that you can include in any class that needs to run code with circuit breaker protection. The concern provides a run_with_circuit method that wraps a code block with circuit breaker functionality, which helps prevent cascading failures and improves system resilience. For more information about the circuit breaker pattern, see:

Use CircuitBreaker

To use the CircuitBreaker concern, you need to include it in a class. For example:

class MyService
  include Gitlab::Llm::Concerns::CircuitBreaker

  def call_external_service
    run_with_circuit do
      # Code that interacts with external service goes here

      raise InternalServerError
    end
  end
end

The call_external_service method is an example method that interacts with an external service. By wrapping the code that interacts with the external service with run_with_circuit, the method is executed within the circuit breaker. The circuit breaker is created and configured by the circuit method, which is called automatically when the CircuitBreaker module is included. The method should raise InternalServerError error which will be counted towards the error threshold if raised during the execution of the code block.

The circuit breaker tracks the number of errors and the rate of requests, and opens the circuit if it reaches the configured error threshold or volume threshold. If the circuit is open, subsequent requests fail fast without executing the code block, and the circuit breaker periodically allows a small number of requests through to test the service's availability before closing the circuit again.

Configuration

The circuit breaker is configured with two constants which control the number of errors and requests at which the circuit will open:

ERROR_THRESHOLD
VOLUME_THRESHOLD

You can adjust these values as needed for the specific service and usage pattern. The InternalServerError is the exception class counted towards the error threshold if raised during the execution of the code block. This is the exception class that triggers the circuit breaker when raised by the code that interacts with the external service.

NOTE: The CircuitBreaker module depends on the Circuitbox gem to provide the circuit breaker implementation. By default, the service name is inferred from the class name where the concern module is included. Override the service_name method if the name needs to be different.

Testing

To test code that uses the CircuitBreaker concern, you can use RSpec shared examples and pass the service and subject variables:

it_behaves_like 'has circuit breaker' do
  let(:service) { dummy_class.new }
  let(:subject) { service.dummy_method }
end

How to implement a new action

Register a new method

Go to the Llm::ExecuteMethodService and add a new method with the new service class you will create.

class ExecuteMethodService < BaseService
  METHODS = {
    # ...
    amazing_new_ai_feature: Llm::AmazingNewAiFeatureService
  }.freeze

Create a Service

Create a new service under ee/app/services/llm/ and inherit it from the BaseService.
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/amazing_new_ai_feature_service.rb

module Llm
  class AmazingNewAiFeatureService < BaseService
    private

    def perform
      ::Llm::CompletionWorker.perform_async(user.id, resource.id, resource.class.name, :amazing_new_ai_feature)
      success
    end

    def valid?
      super && Ability.allowed?(user, :amazing_new_ai_feature, resource)
    end
  end
end

Authorization

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

General AI feature flag is enabled
Feature specific feature flag is enabled
The namespace has the required license for the feature
User is a member of the group/project
experiment_features_enabled and third_party_ai_features_enabled flags are set on the Namespace

For our example, we need to implement the allowed?(:amazing_new_ai_feature) call. As an example, you can look at the Issue Policy for the summarize comments feature. In our example case, we want to implement the feature for Issues as well:

# ee/app/policies/ee/issue_policy.rb

module EE
  module IssuePolicy
    extend ActiveSupport::Concern
    prepended do
      with_scope :subject
      condition(:ai_available) do
        ::Feature.enabled?(:openai_experimentation)
      end

      with_scope :subject
      condition(:amazing_new_ai_feature_enabled) do
        ::Feature.enabled?(:amazing_new_ai_feature, subject_container) &&
          subject_container.licensed_feature_available?(:amazing_new_ai_feature)
      end

      rule do
        ai_available & amazing_new_ai_feature_enabled & is_project_member
      end.enable :amazing_new_ai_feature
    end
  end
end

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 especially useful for chat functionality. For other services, caching is disabled. (It can be enabled for a service by using 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 are two settings allowed on root namespace level that restrict the use of AI features:

experiment_features_enabled
third_party_ai_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 was moved to GA and does not use a third-party setting.
If feature is in GA, but uses a third-party setting, the class will return a proper answer based on the namespace third-party setting.

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 OpenAI and implement two new classes:

# /ee/lib/gitlab/llm/open_ai/completions/amazing_new_ai_feature.rb

module Gitlab
  module Llm
    module OpenAi
      module Completions
        class AmazingNewAiFeature
          def initialize(ai_prompt_class)
            @ai_prompt_class = ai_prompt_class
          end

          def execute(user, issue, options)
            options = ai_prompt_class.get_options(options[:messages])

            ai_response = Gitlab::Llm::OpenAi::Client.new(user).chat(content: nil, **options)

            ::Gitlab::Llm::OpenAi::ResponseService.new(user, issue, ai_response, options: {}).execute(
              Gitlab::Llm::OpenAi::ResponseModifiers::Chat.new
            )
          end

          private

          attr_reader :ai_prompt_class
        end
      end
    end
  end
end

# /ee/lib/gitlab/llm/open_ai/templates/amazing_new_ai_feature.rb

module Gitlab
  module Llm
    module OpenAi
      module Templates
        class AmazingNewAiFeature
          TEMPERATURE = 0.3

          def self.get_options(messages)
            system_content = <<-TEMPLATE
              You are an assistant that writes code for the following input:
              """
            TEMPLATE

            {
              messages: [
                { role: "system", content: system_content },
                { role: "user", content: messages },
              ],
              temperature: TEMPERATURE
            }
          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)
Gitlab::Llm::Anthropic::Client.new(user)

Add Ai Action to GraphQL

TODO

Security

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

25 KiB Raw Blame History

AI features based on 3rd-party integrations

Features

Feature flags

Implement a new AI action

Test AI features locally

Set up the embedding database

Set up GitLab Duo Chat

Tips for local development

Working with GitLab Duo Chat

Setup for GitLab documentation chat (legacy chat)

Contributing to GitLab Duo Chat

Adding a new tool

Debugging

Configure GCP Vertex access

Configure OpenAI access

Configure Anthropic access

Testing GitLab Duo Chat with predefined questions

Experimental REST API

Abstraction layer

GraphQL API

How to receive a response

Current abstraction layer flow

CircuitBreaker

Use CircuitBreaker

Configuration

Testing

How to implement a new action

Register a new method

Create a Service

Authorization

Pairing requests with responses

Caching

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

Implement calls to AI APIs and the prompts

Add Ai Action to GraphQL

Security

25 KiB

Raw Blame History