diff --git a/config/feature_flags/development/summarize_notes_with_anthropic.yml b/config/feature_flags/development/summarize_notes_with_anthropic.yml
deleted file mode 100644
index 145d75e2781560238c58a7a45b5a4bdb962e4cae..0000000000000000000000000000000000000000
--- a/config/feature_flags/development/summarize_notes_with_anthropic.yml
+++ /dev/null
@@ -1,8 +0,0 @@
----
-name: summarize_notes_with_anthropic
-introduced_by_url: https://gitlab.com/gitlab-org/gitlab/-/merge_requests/134731
-rollout_issue_url: https://gitlab.com/gitlab-org/gitlab/-/work_items/430196
-milestone: '16.6'
-type: development
-group: group::duo chat
-default_enabled: true
diff --git a/ee/lib/gitlab/llm/anthropic/client.rb b/ee/lib/gitlab/llm/anthropic/client.rb
index 74b97363d540f0879a6af53dd86302c91dba0682..73805ab1ba1923bc3eeeaf2cf81242989cd20b50 100644
--- a/ee/lib/gitlab/llm/anthropic/client.rb
+++ b/ee/lib/gitlab/llm/anthropic/client.rb
@@ -76,6 +76,24 @@ def messages_complete(messages:, **options)
response
end
+ def messages_stream(messages:, **options)
+ return unless enabled?
+
+ response_body = ""
+ perform_messages_request(messages: messages, options: options.merge(stream: true)) do |parsed_event|
+ response_body += parsed_event.dig('delta', 'text') if parsed_event.dig('delta', 'text')
+
+ yield parsed_event if block_given?
+ end
+ logger.info_or_debug(user, message: "Received response from Anthropic", response: response_body)
+
+ track_prompt_size(token_size(messages))
+ track_response_size(token_size(response_body))
+
+ response_body
+ end
+ traceable :messages_stream, name: 'Request to Anthropic', run_type: 'llm'
+
private
attr_reader :user, :logger, :tracking_context, :unit_primitive
@@ -109,7 +127,11 @@ def perform_messages_request(messages:, options:)
timeout: timeout,
allow_local_requests: true,
stream_body: options.fetch(:stream, false)
- )
+ ) do |fragment|
+ parse_sse_events(fragment).each do |parsed_event|
+ yield parsed_event if block_given?
+ end
+ end
raise Gitlab::AiGateway::ForbiddenError if response.forbidden?
diff --git a/ee/lib/gitlab/llm/chain/requests/anthropic.rb b/ee/lib/gitlab/llm/chain/requests/anthropic.rb
index 1d65091ed74dde5046cddcd56b10b47220a3b929..bbf274d17287628114e6eb5089f3cdc8cd9825e3 100644
--- a/ee/lib/gitlab/llm/chain/requests/anthropic.rb
+++ b/ee/lib/gitlab/llm/chain/requests/anthropic.rb
@@ -7,8 +7,6 @@ module Requests
class Anthropic < Base
attr_reader :ai_client
- TEMPERATURE = 0.1
- STOP_WORDS = ["\n\nHuman", "Observation:"].freeze
PROMPT_SIZE = 30_000
def initialize(user, unit_primitive:, tracking_context: {})
@@ -20,15 +18,14 @@ def initialize(user, unit_primitive:, tracking_context: {})
# TODO: unit primitive param is temporarily added to provide parity with ai_gateway-related method
def request(prompt, unit_primitive: nil) # rubocop: disable Lint/UnusedMethodArgument -- added to provide parity with ai_gateway-related method
- return unless prompt[:prompt]
+ return unless prompt[:messages]
- ai_client.stream(
- prompt: prompt[:prompt],
- **default_options.merge(prompt.fetch(:options, {}))
+ ai_client.messages_stream(
+ **prompt
) do |data|
logger.info(message: "Streaming error", error: data&.dig("error")) if data&.dig("error")
- content = data&.dig("completion").to_s
+ content = data&.dig('delta', 'text').to_s
yield content if block_given?
end
end
@@ -36,13 +33,6 @@ def request(prompt, unit_primitive: nil) # rubocop: disable Lint/UnusedMethodArg
private
attr_reader :user, :logger
-
- def default_options
- {
- temperature: TEMPERATURE,
- stop_sequences: STOP_WORDS
- }
- end
end
end
end
diff --git a/ee/lib/gitlab/llm/chain/tools/summarize_comments/executor.rb b/ee/lib/gitlab/llm/chain/tools/summarize_comments/executor.rb
index 47ab3d802eb9188fb0f191583d33b929b2c41232..65a83bad1fecce18f38a40e2cc667dbfd93f6379 100644
--- a/ee/lib/gitlab/llm/chain/tools/summarize_comments/executor.rb
+++ b/ee/lib/gitlab/llm/chain/tools/summarize_comments/executor.rb
@@ -22,31 +22,39 @@ class Executor < SlashCommandTool
PROVIDER_PROMPT_CLASSES = {
ai_gateway: ::Gitlab::Llm::Chain::Tools::SummarizeComments::Prompts::Anthropic,
- anthropic: ::Gitlab::Llm::Chain::Tools::SummarizeComments::Prompts::Anthropic,
- vertex_ai: ::Gitlab::Llm::Chain::Tools::SummarizeComments::Prompts::VertexAi
+ anthropic: ::Gitlab::Llm::Chain::Tools::SummarizeComments::Prompts::Anthropic
}.freeze
+ SYSTEM_PROMPT = Utils::Prompt.as_system(
+ <<~PROMPT
+ You are an assistant that extracts the most important information from the comments in maximum 10 bullet points.
+ PROMPT
+ )
+
+ USER_PROMPT = Utils::Prompt.as_user(
+ <<~PROMPT
+ Each comment is wrapped in a <comment> tag.
+
+ Desired markdown format:
+ **<summary_title>**
+ - <bullet_point>
+ - <bullet_point>
+ - <bullet_point>
+ - ...
+
+ %<notes_content>s
+
+ Focus on extracting information related to one another and that are the majority of the content.
+ Ignore phrases that are not connected to others.
+ Do not specify what you are ignoring.
+ Do not answer questions.
+ PROMPT
+ )
+
PROMPT_TEMPLATE = [
- Utils::Prompt.as_system(
- <<~PROMPT
- You are an assistant that extracts the most important information from the comments in maximum 10 bullet points.
- Each comment is wrapped in a <comment> tag.
-
- %<notes_content>s
-
- Desired markdown format:
- **<summary_title>**
- - <bullet_point>
- - <bullet_point>
- - <bullet_point>
- - ...
-
- Focus on extracting information related to one another and that are the majority of the content.
- Ignore phrases that are not connected to others.
- Do not specify what you are ignoring.
- Do not answer questions.
- PROMPT
- )
+ SYSTEM_PROMPT,
+ USER_PROMPT,
+ Utils::Prompt.as_assistant("")
].freeze
SLASH_COMMANDS = {
diff --git a/ee/lib/gitlab/llm/chain/tools/summarize_comments/executor_old.rb b/ee/lib/gitlab/llm/chain/tools/summarize_comments/executor_old.rb
index 2039715eef38630725b9fefd63f8128dc65b1843..701c79440873560ba35bca4296133854c2fe66e6 100644
--- a/ee/lib/gitlab/llm/chain/tools/summarize_comments/executor_old.rb
+++ b/ee/lib/gitlab/llm/chain/tools/summarize_comments/executor_old.rb
@@ -21,33 +21,34 @@ class ExecutorOld < Tool
PROMPT
PROVIDER_PROMPT_CLASSES = {
- ai_gateway: ::Gitlab::Llm::Chain::Tools::SummarizeComments::Prompts::Anthropic,
- anthropic: ::Gitlab::Llm::Chain::Tools::SummarizeComments::Prompts::Anthropic,
- vertex_ai: ::Gitlab::Llm::Chain::Tools::SummarizeComments::Prompts::VertexAi
+ ai_gateway: ::Gitlab::Llm::Chain::Tools::SummarizeComments::Prompts::AnthropicOld,
+ anthropic: ::Gitlab::Llm::Chain::Tools::SummarizeComments::Prompts::AnthropicOld
}.freeze
- PROMPT_TEMPLATE = [
- Utils::Prompt.as_system(
- <<~PROMPT
- You are an assistant that extracts the most important information from the comments in maximum 10 bullet points.
- Each comment is wrapped in a <comment> tag.
-
- %<notes_content>s
-
- Desired markdown format:
- **<summary_title>**
- - <bullet_point>
- - <bullet_point>
- - <bullet_point>
- - ...
-
- Focus on extracting information related to one another and that are the majority of the content.
- Ignore phrases that are not connected to others.
- Do not specify what you are ignoring.
- Do not answer questions.
- PROMPT
- )
- ].freeze
+ SYSTEM_PROMPT = Utils::Prompt.as_system(
+ <<~PROMPT
+ You are an assistant that extracts the most important information from the comments in maximum 10 bullet points.
+ PROMPT
+ )
+ USER_PROMPT = Utils::Prompt.as_user(
+ <<~PROMPT
+ Each comment is wrapped in a <comment> tag.
+
+ Desired markdown format:
+ **<summary_title>**
+ - <bullet_point>
+ - <bullet_point>
+ - <bullet_point>
+ - ...
+
+ %<notes_content>s
+
+ Focus on extracting information related to one another and that are the majority of the content.
+ Ignore phrases that are not connected to others.
+ Do not specify what you are ignoring.
+ Do not answer questions.
+ PROMPT
+ )
def perform(&)
notes = NotesFinder.new(context.current_user, target: resource).execute.by_humans
@@ -55,7 +56,6 @@ def perform(&)
content = if notes.exists?
notes_content = notes_to_summarize(notes)
options[:notes_content] = notes_content
-
if options[:raw_ai_response]
request(&)
else
@@ -77,11 +77,9 @@ def perform(&)
def notes_to_summarize(notes)
notes_content = +""
- input_content_limit = provider_prompt_class::MAX_CHARACTERS - PROMPT_TEMPLATE.size
+ input_content_limit = provider_prompt_class::MAX_CHARACTERS - SYSTEM_PROMPT.size - USER_PROMPT.size
notes.each_batch do |batch|
batch.pluck(:id, :note).each do |note| # rubocop: disable CodeReuse/ActiveRecord -- we need to pluck just id and note
- input_content_limit = provider_prompt_class::MAX_CHARACTERS
-
break notes_content if notes_content.size + note[1].size >= input_content_limit
notes_content << (format("<comment>%<note>s</comment>", note: note[1]))
diff --git a/ee/lib/gitlab/llm/chain/tools/summarize_comments/prompts/anthropic.rb b/ee/lib/gitlab/llm/chain/tools/summarize_comments/prompts/anthropic.rb
index 6b7d66f10a5251012d11818e1942cab8fcda8397..0c3182c62268f24de5c2039e427a9c8962d7aff3 100644
--- a/ee/lib/gitlab/llm/chain/tools/summarize_comments/prompts/anthropic.rb
+++ b/ee/lib/gitlab/llm/chain/tools/summarize_comments/prompts/anthropic.rb
@@ -9,20 +9,18 @@ module Prompts
class Anthropic
include Concerns::AnthropicPrompt
- OUTPUT_TOKEN_LIMIT = 2048
-
- def self.prompt(options)
- base_prompt = Utils::Prompt.no_role_text(
- ::Gitlab::Llm::Chain::Tools::SummarizeComments::Executor::PROMPT_TEMPLATE, options
+ def self.prompt(variables)
+ conversation = Utils::Prompt.role_conversation(
+ Utils::Prompt.format_conversation(
+ ::Gitlab::Llm::Chain::Tools::SummarizeComments::Executor::PROMPT_TEMPLATE,
+ variables
+ )
)
- Requests::Anthropic.prompt(
- "\n\nHuman: #{base_prompt}\n\nAssistant:",
- options: {
- model: ::Gitlab::Llm::Anthropic::Client::DEFAULT_INSTANT_MODEL,
- max_tokens_to_sample: OUTPUT_TOKEN_LIMIT
- }
- )
+ {
+ prompt: conversation,
+ options: { model: ::Gitlab::Llm::Anthropic::Client::CLAUDE_3_5_SONNET }
+ }
end
end
end
diff --git a/ee/lib/gitlab/llm/chain/tools/summarize_comments/prompts/anthropic_old.rb b/ee/lib/gitlab/llm/chain/tools/summarize_comments/prompts/anthropic_old.rb
new file mode 100644
index 0000000000000000000000000000000000000000..72dbfecb81779928227dd9327978ecea4413cf15
--- /dev/null
+++ b/ee/lib/gitlab/llm/chain/tools/summarize_comments/prompts/anthropic_old.rb
@@ -0,0 +1,34 @@
+# frozen_string_literal: true
+
+module Gitlab
+ module Llm
+ module Chain
+ module Tools
+ module SummarizeComments
+ module Prompts
+ class AnthropicOld
+ include Concerns::AnthropicPrompt
+
+ TEMPERATURE = 0.1
+
+ def self.prompt(variables)
+ {
+ messages: Gitlab::Llm::Chain::Utils::Prompt.role_conversation(
+ Gitlab::Llm::Chain::Utils::Prompt.format_conversation([
+ ::Gitlab::Llm::Chain::Tools::SummarizeComments::ExecutorOld::USER_PROMPT,
+ Utils::Prompt.as_assistant("")
+ ], variables)
+ ),
+ system: Gitlab::Llm::Chain::Utils::Prompt.no_role_text(
+ [::Gitlab::Llm::Chain::Tools::SummarizeComments::ExecutorOld::SYSTEM_PROMPT], variables),
+ model: ::Gitlab::Llm::Anthropic::Client::CLAUDE_3_5_SONNET,
+ temperature: TEMPERATURE
+ }
+ end
+ end
+ end
+ end
+ end
+ end
+ end
+end
diff --git a/ee/lib/gitlab/llm/chain/tools/summarize_comments/prompts/vertex_ai.rb b/ee/lib/gitlab/llm/chain/tools/summarize_comments/prompts/vertex_ai.rb
deleted file mode 100644
index 175f75c38fc5cda2b6a57acc731ff0dd3e25babf..0000000000000000000000000000000000000000
--- a/ee/lib/gitlab/llm/chain/tools/summarize_comments/prompts/vertex_ai.rb
+++ /dev/null
@@ -1,32 +0,0 @@
-# frozen_string_literal: true
-
-module Gitlab
- module Llm
- module Chain
- module Tools
- module SummarizeComments
- module Prompts
- class VertexAi
- include Concerns::VertexAiPrompt
-
- OUTPUT_TOKEN_LIMIT = 1024
-
- def self.prompt(options)
- prompt = Utils::Prompt.no_role_text(
- ::Gitlab::Llm::Chain::Tools::SummarizeComments::Executor::PROMPT_TEMPLATE, options
- )
-
- {
- prompt: prompt,
- options: {
- max_output_tokens: OUTPUT_TOKEN_LIMIT
- }
- }
- end
- end
- end
- end
- end
- end
- end
-end
diff --git a/ee/lib/gitlab/llm/completions/summarize_all_open_notes.rb b/ee/lib/gitlab/llm/completions/summarize_all_open_notes.rb
index e31ccf3dc01469327c11a9afd9644ee177b6794f..7a49b5613832a772b72a17e87d150eb3554e4194 100644
--- a/ee/lib/gitlab/llm/completions/summarize_all_open_notes.rb
+++ b/ee/lib/gitlab/llm/completions/summarize_all_open_notes.rb
@@ -37,13 +37,8 @@ def execute
private
def ai_provider_request(user)
- if Feature.enabled?(:summarize_notes_with_anthropic, user)
- ::Gitlab::Llm::Chain::Requests::Anthropic.new(user,
- unit_primitive: 'summarize_issue_discussions', tracking_context: tracking_context)
- else
- ::Gitlab::Llm::Chain::Requests::VertexAi.new(user,
- unit_primitive: 'summarize_issue_discussions', tracking_context: tracking_context)
- end
+ ::Gitlab::Llm::Chain::Requests::Anthropic.new(user,
+ unit_primitive: 'summarize_issue_discussions', tracking_context: tracking_context)
end
def issuable
diff --git a/ee/spec/lib/gitlab/llm/chain/concerns/ai_dependent_spec.rb b/ee/spec/lib/gitlab/llm/chain/concerns/ai_dependent_spec.rb
index f07d680f54ffd06c121a90cf9338b184961d06f5..4bfd32c7690ff047dfdaa59282e1d35bf426c64c 100644
--- a/ee/spec/lib/gitlab/llm/chain/concerns/ai_dependent_spec.rb
+++ b/ee/spec/lib/gitlab/llm/chain/concerns/ai_dependent_spec.rb
@@ -27,6 +27,36 @@
tool.prompt
end
+ context "when calling summarize comments tool" do
+ let_it_be(:project) { create(:project) }
+ let_it_be(:issue) { create(:issue, project: project) }
+ let_it_be(:note) do
+ create(:note_on_issue, noteable: issue, project: project, note: "Please correct this small nit")
+ end
+
+ let(:context) do
+ Gitlab::Llm::Chain::GitlabContext.new(
+ current_user: user, container: double, resource: issue, ai_request: ai_request
+ )
+ end
+
+ let(:tool) do
+ ::Gitlab::Llm::Chain::Tools::SummarizeComments::Executor.new(
+ context: context,
+ options: {
+ input: 'Summarize issue comments.',
+ notes_content: "<comment>#{note.note}</comment>"
+ }
+ )
+ end
+
+ it "returns prompt" do
+ expect(tool.class::PROVIDER_PROMPT_CLASSES[:anthropic]).to receive(:prompt).and_call_original
+
+ tool.prompt
+ end
+ end
+
context 'when there are no provider prompt classes' do
let(:dummy_tool_class) do
Class.new(::Gitlab::Llm::Chain::Tools::Tool) do
diff --git a/ee/spec/lib/gitlab/llm/chain/requests/anthropic_spec.rb b/ee/spec/lib/gitlab/llm/chain/requests/anthropic_spec.rb
index 7de998482506f2e0b54543fd4027f2b9c7b19fcd..731d99f9b52b086ae3e962d5e64bb6f3a048659d 100644
--- a/ee/spec/lib/gitlab/llm/chain/requests/anthropic_spec.rb
+++ b/ee/spec/lib/gitlab/llm/chain/requests/anthropic_spec.rb
@@ -19,12 +19,27 @@
let(:instance) { described_class.new(user, unit_primitive: 'duo_chat') }
let(:logger) { instance_double(Gitlab::Llm::Logger) }
let(:ai_client) { double }
- let(:response) { { "completion" => "Hello World " } }
+ let(:prompt_message) do
+ [
+ {
+ role: :user,
+ content: "Some user request"
+ }
+ ]
+ end
+
+ let(:response) do
+ {
+ "delta" => {
+ "type" => "text_delta",
+ "text" => "Hello World"
+ }
+ }
+ end
+
let(:expected_params) do
{
- prompt: "some user request",
- temperature: 0.1,
- stop_sequences: ["\n\nHuman", "Observation:"]
+ messages: prompt_message
}
end
@@ -33,41 +48,41 @@
allow(instance).to receive(:ai_client).and_return(ai_client)
end
- context 'with prompt and options' do
- let(:params) { { prompt: "some user request", options: { max_tokens: 4000 } } }
+ context 'with prompt' do
+ let(:params) do
+ { messages: prompt_message, max_tokens: 4000 }
+ end
- it 'calls the anthropic streaming endpoint and yields response without stripping it' do
- expect(ai_client).to receive(:stream).with(expected_params.merge({ max_tokens: 4000 })).and_yield(response)
+ it 'calls the anthropic messages streaming endpoint and yields response without stripping it' do
+ expect(ai_client).to receive(:messages_stream).with(expected_params.merge(max_tokens: 4000)).and_yield(response)
- expect { |b| instance.request(params, &b) }.to yield_with_args(
- "Hello World "
- )
+ expect { |b| instance.request(params, &b) }.to yield_with_args("Hello World")
end
it 'returns the response from anthropic' do
- expect(ai_client).to receive(:stream).with(expected_params.merge({ max_tokens: 4000 }))
- .and_return(response["completion"])
+ expect(ai_client).to receive(:messages_stream).with(expected_params.merge({ max_tokens: 4000 }))
+ .and_return(response)
- expect(request).to eq("Hello World ")
+ expect(request["delta"]["text"]).to eq("Hello World")
end
end
context 'when options are not present' do
- let(:params) { { prompt: "some user request" } }
+ let(:params) { { messages: prompt_message } }
it 'calls the anthropic streaming endpoint' do
- expect(ai_client).to receive(:stream).with(expected_params)
+ expect(ai_client).to receive(:messages_stream).with(expected_params)
request
end
end
context 'when stream errors' do
- let(:params) { { prompt: "some user request" } }
+ let(:params) { { messages: prompt_message } }
let(:response) { { "error" => { "type" => "overload_error", message: "Overloaded" } } }
it 'logs the error' do
- expect(ai_client).to receive(:stream).with(expected_params).and_yield(response)
+ expect(ai_client).to receive(:messages_stream).with(expected_params).and_yield(response)
expect(logger).to receive(:info).with(hash_including(message: "Streaming error", error: response["error"]))
request
diff --git a/ee/spec/lib/gitlab/llm/chain/tools/summarize_comments/executor_spec.rb b/ee/spec/lib/gitlab/llm/chain/tools/summarize_comments/executor_spec.rb
index 617405e3c9d49bb522f4e24e5886c3cbb8c49c79..12a69f18449b6c732dd70cf1df843cd5d7712177 100644
--- a/ee/spec/lib/gitlab/llm/chain/tools/summarize_comments/executor_spec.rb
+++ b/ee/spec/lib/gitlab/llm/chain/tools/summarize_comments/executor_spec.rb
@@ -55,11 +55,12 @@
expected_prompt = <<~PROMPT.chomp
You are an assistant that extracts the most important information from the comments in maximum 10 bullet points.
- Each comment is wrapped in a <comment> tag.
PROMPT
- expect(prompt).to include(expected_prompt)
- expect(prompt).to include(note.note)
+ system_prompt = prompt[0][:content]
+ user_prompt = prompt[1][:content]
+ expect(system_prompt).to include(expected_prompt)
+ expect(user_prompt).to include(note.note)
end
it 'sets the correct unit primitive' do
diff --git a/ee/spec/lib/gitlab/llm/chain/tools/summarize_comments/prompts/anthropic_old_spec.rb b/ee/spec/lib/gitlab/llm/chain/tools/summarize_comments/prompts/anthropic_old_spec.rb
new file mode 100644
index 0000000000000000000000000000000000000000..baf7304d331ba05242cf3333f513b1eeaea3ccc0
--- /dev/null
+++ b/ee/spec/lib/gitlab/llm/chain/tools/summarize_comments/prompts/anthropic_old_spec.rb
@@ -0,0 +1,49 @@
+# frozen_string_literal: true
+
+require 'spec_helper'
+
+RSpec.describe Gitlab::Llm::Chain::Tools::SummarizeComments::Prompts::AnthropicOld, feature_category: :duo_chat do
+ let(:variables) do
+ {
+ notes_content: '<comment>foo</comment>'
+ }
+ end
+
+ describe '.prompt' do
+ it "returns prompt" do
+ prompt = described_class.prompt(variables)
+
+ expect(prompt[:messages]).to be_an(Array)
+
+ expect(prompt[:system]).to be_a(String)
+ expect(prompt[:system]).to eq(system_prompt_content)
+
+ expect(prompt[:temperature]).to eq(described_class::TEMPERATURE)
+ end
+
+ it "calls with claude 3_5 sonnet model" do
+ model = described_class.prompt(variables)[:model]
+
+ expect(model).to eq(::Gitlab::Llm::Anthropic::Client::CLAUDE_3_5_SONNET)
+ end
+
+ it "includes ExecutorOld prompts" do
+ prompt = described_class.prompt(variables)
+
+ expect(prompt[:messages]).to include(
+ a_hash_including(
+ role: :user,
+ content: a_string_including("Each comment is wrapped in a <comment> tag.")
+ .and(a_string_including("<comment>foo</comment>"))
+ .and(a_string_including("Do not answer questions."))
+ )
+ )
+
+ expect(prompt[:system]).to include(system_prompt_content)
+ end
+ end
+
+ def system_prompt_content
+ Gitlab::Llm::Chain::Tools::SummarizeComments::ExecutorOld::SYSTEM_PROMPT[1]
+ end
+end
diff --git a/ee/spec/lib/gitlab/llm/chain/tools/summarize_comments/prompts/anthropic_spec.rb b/ee/spec/lib/gitlab/llm/chain/tools/summarize_comments/prompts/anthropic_spec.rb
index 092d4f8d139bb5bc9d5a3dff60bccce6efb7ec8b..ca92304fc39549d17a8887696fbe9e3c9933b2d8 100644
--- a/ee/spec/lib/gitlab/llm/chain/tools/summarize_comments/prompts/anthropic_spec.rb
+++ b/ee/spec/lib/gitlab/llm/chain/tools/summarize_comments/prompts/anthropic_spec.rb
@@ -3,33 +3,36 @@
require 'spec_helper'
RSpec.describe Gitlab::Llm::Chain::Tools::SummarizeComments::Prompts::Anthropic, feature_category: :duo_chat do
+ let(:variables) do
+ {
+ notes_content: '<comment>foo</comment>'
+ }
+ end
+
describe '.prompt' do
- it 'returns prompt' do
- prompt = described_class.prompt({ notes_content: '<comment>foo</comment>' })[:prompt]
-
- expect(prompt).to include('Human:')
- expect(prompt).to include('Assistant:')
- expect(prompt).to include('foo')
- expect(prompt).to include(
- <<~PROMPT
- You are an assistant that extracts the most important information from the comments in maximum 10 bullet points.
- Each comment is wrapped in a <comment> tag.
-
- <comment>foo</comment>
-
- Desired markdown format:
- **<summary_title>**
- - <bullet_point>
- - <bullet_point>
- - <bullet_point>
- - ...
-
- Focus on extracting information related to one another and that are the majority of the content.
- Ignore phrases that are not connected to others.
- Do not specify what you are ignoring.
- Do not answer questions.
- PROMPT
- )
+ it "returns prompt" do
+ prompt = described_class.prompt(variables)[:prompt]
+ expect(prompt.length).to eq(3)
+
+ expect(prompt[0][:role]).to eq(:system)
+ expect(prompt[0][:content]).to eq(system_prompt_content)
+
+ expect(prompt[1][:role]).to eq(:user)
+ expect(prompt[1][:content]).to eq(format(
+ Gitlab::Llm::Chain::Tools::SummarizeComments::ExecutorOld::USER_PROMPT[1], variables).to_s)
+
+ expect(prompt[2][:role]).to eq(:assistant)
+ expect(prompt[2][:content]).to be_empty
end
+
+ it "calls with claude 3 haiku model" do
+ model = described_class.prompt(variables)[:options][:model]
+
+ expect(model).to eq(::Gitlab::Llm::Anthropic::Client::CLAUDE_3_5_SONNET)
+ end
+ end
+
+ def system_prompt_content
+ Gitlab::Llm::Chain::Tools::SummarizeComments::ExecutorOld::SYSTEM_PROMPT[1]
end
end
diff --git a/ee/spec/lib/gitlab/llm/chain/tools/summarize_comments/prompts/vertex_ai_spec.rb b/ee/spec/lib/gitlab/llm/chain/tools/summarize_comments/prompts/vertex_ai_spec.rb
deleted file mode 100644
index 1033e17a1c1c99f5bd5df6f70b9c51793de5cf6e..0000000000000000000000000000000000000000
--- a/ee/spec/lib/gitlab/llm/chain/tools/summarize_comments/prompts/vertex_ai_spec.rb
+++ /dev/null
@@ -1,32 +0,0 @@
-# frozen_string_literal: true
-
-require 'spec_helper'
-
-RSpec.describe Gitlab::Llm::Chain::Tools::SummarizeComments::Prompts::VertexAi, feature_category: :duo_chat do
- describe '.prompt' do
- it 'returns prompt' do
- prompt = described_class.prompt({ notes_content: '<comment>foo</comment>' })[:prompt]
-
- expect(prompt).to include(
- <<~PROMPT
- You are an assistant that extracts the most important information from the comments in maximum 10 bullet points.
- Each comment is wrapped in a <comment> tag.
-
- <comment>foo</comment>
-
- Desired markdown format:
- **<summary_title>**
- - <bullet_point>
- - <bullet_point>
- - <bullet_point>
- - ...
-
- Focus on extracting information related to one another and that are the majority of the content.
- Ignore phrases that are not connected to others.
- Do not specify what you are ignoring.
- Do not answer questions.
- PROMPT
- )
- end
- end
-end
diff --git a/ee/spec/lib/gitlab/llm/completions/summarize_all_open_notes_spec.rb b/ee/spec/lib/gitlab/llm/completions/summarize_all_open_notes_spec.rb
index 638172f6a1f1bc4470e4ab9c0b2f1e46476c4a2c..f4bd3961a4eeac32af081191a515ca2fe37db6e2 100644
--- a/ee/spec/lib/gitlab/llm/completions/summarize_all_open_notes_spec.rb
+++ b/ee/spec/lib/gitlab/llm/completions/summarize_all_open_notes_spec.rb
@@ -20,10 +20,10 @@
)
end
- RSpec.shared_examples 'performs completion' do
+ RSpec.shared_examples 'performs messages stream' do
it 'returns summary' do
expect_next_instance_of(ai_request_class) do |instance|
- expect(instance).to receive(completion_method).and_return(ai_response)
+ expect(instance).to receive(messages_method).and_return(ai_response)
end
response_modifier = double
@@ -109,7 +109,7 @@
describe "#execute", :saas do
let(:ai_request_class) { ::Gitlab::Llm::Anthropic::Client }
- let(:completion_method) { :stream }
+ let(:messages_method) { :messages_stream }
let(:options) { {} }
let_it_be(:user) { create(:user) }
@@ -149,20 +149,7 @@
let_it_be(:notes) { create_pair(:note_on_issue, project: project, noteable: issuable) }
let_it_be(:system_note) { create(:note_on_issue, :system, project: project, noteable: issuable) }
- # anthropic as provider as summarize_notes_with_anthropic is enabled by default.
- it_behaves_like 'performs completion'
-
- context 'with vertex_ai provider' do
- let(:completion_method) { :text }
- let(:ai_request_class) { ::Gitlab::Llm::VertexAi::Client }
- let(:ai_response) { { "predictions" => [{ "content" => "some ai response text" }] } }
-
- before do
- stub_feature_flags(summarize_notes_with_anthropic: false)
- end
-
- it_behaves_like 'performs completion'
- end
+ it_behaves_like 'performs messages stream'
end
context 'for a work item' do
@@ -170,7 +157,7 @@
let_it_be(:notes) { create_pair(:note_on_work_item, project: project, noteable: issuable) }
let_it_be(:system_note) { create(:note_on_work_item, :system, project: project, noteable: issuable) }
- it_behaves_like 'performs completion'
+ it_behaves_like 'performs messages stream'
end
context 'for a merge request' do
@@ -187,7 +174,7 @@
let_it_be(:notes) { create_pair(:note_on_epic, noteable: issuable) }
let_it_be(:system_note) { create(:note_on_epic, :system, noteable: issuable) }
- it_behaves_like 'performs completion'
+ it_behaves_like 'performs messages stream'
end
end
end