FEATURE: Prompts can consist of multiple messages. (#21)

A prompt with multiple messages leads to better results, as the AI can learn for given examples. Alongside this change, we provide a better default proofreading prompt.

app/models/completion_prompt.rb

@@ -1,7 +1,38 @@
 # frozen_string_literal: true
 
 class CompletionPrompt < ActiveRecord::Base
+  # TODO(roman): Remove sept 2023.
+  self.ignored_columns = ["value"]
+
+  VALID_ROLES = %w[system user assistant]
+
   enum :prompt_type, { text: 0, list: 1, diff: 2 }
+
+  validates :messages, length: { maximum: 20 }
+  validate :each_message_length
+  validate :each_message_role
+
+  def messages_with_user_input(user_input)
+    self.messages << { role: "user", content: user_input }
+  end
+
+  private
+
+  def each_message_length
+    messages.each_with_index do |msg, idx|
+      next if msg["content"].length <= 1000
+
+      errors.add(:messages, I18n.t("errors.prompt_message_length", idx: idx + 1))
+    end
+  end
+
+  def each_message_role
+    messages.each_with_index do |msg, idx|
+      next if VALID_ROLES.include?(msg["role"])
+
+      errors.add(:messages, I18n.t("errors.invalid_prompt_role", idx: idx + 1))
+    end
+  end
 end
 
 # == Schema Information
@@ -12,10 +43,10 @@ end
 #  name            :string           not null
 #  translated_name :string
 #  prompt_type     :integer          default("text"), not null
-#  value           :text             not null
 #  enabled         :boolean          default(TRUE), not null
 #  created_at      :datetime         not null
 #  updated_at      :datetime         not null
+#  messages        :jsonb            not null
 #
 # Indexes
 #

config/locales/server.en.yml

@@ -51,6 +51,9 @@ en:
         flagged_by_toxicity: The AI plugin flagged this after classifying it as toxic.
         flagged_by_nsfw: The AI plugin flagged this after classifying at least one of the attached images as NSFW.
   
+  errors:
+    prompt_message_length: The message %{idx} is over the 1000 character limit.
+    invalid_prompt_role: The message %{idx} has an invalid role.
 
   discourse_ai:
     ai_helper:

db/fixtures/ai-helper/600_completion_prompts.rb

@@ -3,35 +3,80 @@ CompletionPrompt.seed do |cp|
   cp.id = -1
   cp.name = "translate"
   cp.prompt_type = CompletionPrompt.prompt_types[:text]
-  cp.value = <<~STRING
+  cp.messages = [{ role: "system", content: <<~TEXT }]
     I want you to act as an English translator, spelling corrector and improver. I will speak to you
     in any language and you will detect the language, translate it and answer in the corrected and 
     improved version of my text, in English. I want you to replace my simplified A0-level words and 
     sentences with more beautiful and elegant, upper level English words and sentences. 
     Keep the meaning same, but make them more literary. I want you to only reply the correction, 
     the improvements and nothing else, do not write explanations.
-  STRING
+  TEXT
 end
 
 CompletionPrompt.seed do |cp|
   cp.id = -2
   cp.name = "generate_titles"
   cp.prompt_type = CompletionPrompt.prompt_types[:list]
-  cp.value = <<~STRING
+  cp.messages = [{ role: "system", content: <<~TEXT }]
     I want you to act as a title generator for written pieces. I will provide you with a text, 
     and you will generate five attention-grabbing titles. Please keep the title concise and under 20 words,
     and ensure that the meaning is maintained. Replies will utilize the language type of the topic.
     I want you to only reply the list of options and nothing else, do not write explanations.
-  STRING
+  TEXT
 end
 
 CompletionPrompt.seed do |cp|
   cp.id = -3
   cp.name = "proofread"
   cp.prompt_type = CompletionPrompt.prompt_types[:diff]
-  cp.value = <<~STRING
-    I want you act as a proofreader. I will provide you with a text and I want you to review them for any spelling, 
-    grammar, or punctuation errors. Once you have finished reviewing the text, provide me with any necessary 
-    corrections or suggestions for improve the text.
-  STRING
+  cp.messages = [
+    { role: "system", content: <<~TEXT },
+      You are a markdown proofreader. You correct egregious typos and phrasing issues but keep the user's original voice.
+      You do not touch code blocks. I will provide you with text to proofread. If nothing needs fixing, then you will echo the text back.
+
+      Optionally, a user can specify intensity. Intensity 10 is a pedantic English teacher correcting the text.
+      Intensity 1 is a minimal proofreader. By default, you operate at intensity 1.
+    TEXT
+    { role: "user", content: "![amazing car|100x100, 22%](upload://hapy.png)" },
+    { role: "assistant", content: "![Amazing car|100x100, 22%](upload://hapy.png)" },
+    { role: "user", content: <<~TEXT },
+      Intensity 1:
+      The rain in spain stays mainly in the plane.
+    TEXT
+    { role: "assistant", content: "The rain in Spain, stays mainly in the Plane." },
+    { role: "user", content: "The rain in Spain, stays mainly in the Plane." },
+    { role: "assistant", content: "The rain in Spain, stays mainly in the Plane." },
+    { role: "user", content: <<~TEXT },
+      Intensity 1:
+      Hello,
+
+      Sometimes the logo isn't changing automatically when color scheme changes.
+
+      ![Screen Recording 2023-03-17 at 18.04.22|video](upload://2rcVL0ZMxHPNtPWQbZjwufKpWVU.mov)
+    TEXT
+    { role: "assistant", content: <<~TEXT },
+      Hello,
+      Sometimes the logo does not change automatically when the color scheme changes.
+      ![Screen Recording 2023-03-17 at 18.04.22|video](upload://2rcVL0ZMxHPNtPWQbZjwufKpWVU.mov)
+    TEXT
+    { role: "user", content: <<~TEXT },
+      Intensity 1:
+      Any ideas what is wrong with this peace of cod?
+      > This quot contains a typo
+      ```ruby
+      # this has speling mistakes
+      testin.atypo = 11
+      baad = "bad"
+      ```
+    TEXT
+    { role: "assistant", content: <<~TEXT },
+      Any ideas what is wrong with this piece of code?
+      > This quot contains a typo
+      ```ruby
+      # This has spelling mistakes
+      testing.a_typo = 11
+      bad = "bad"
+      ```
+    TEXT
+  ]
 end

db/migrate/20230320185619_multi_message_completion_prompts.rb

@@ -0,0 +1,7 @@
+# frozen_string_literal: true
+
+class MultiMessageCompletionPrompts < ActiveRecord::Migration[7.0]
+  def change
+    add_column :completion_prompts, :messages, :jsonb, null: false
+  end
+end

db/post_migrate/20230320191928_drop_completion_prompt_value.rb

@@ -0,0 +1,7 @@
+# frozen_string_literal: true
+
+class DropCompletionPromptValue < ActiveRecord::Migration[7.0]
+  def change
+    remove_column :completion_prompts, :value, :text
+  end
+end

lib/modules/ai_helper/open_ai_prompt.rb

@@ -23,10 +23,10 @@ module DiscourseAi
       def generate_and_send_prompt(prompt, text)
         result = { type: prompt.prompt_type }
 
-        ai_messages = [{ role: "system", content: prompt.value }, { role: "user", content: text }]
+        messages = prompt.messages_with_user_input(text)
 
         result[:suggestions] = DiscourseAi::Inference::OpenAiCompletions
-          .perform!(ai_messages)
+          .perform!(messages)
           .dig(:choices)
           .to_a
           .flat_map { |choice| parse_content(prompt, choice.dig(:message, :content).to_s) }

lib/shared/inference/openai_completions.rb

@@ -3,7 +3,7 @@
 module ::DiscourseAi
   module Inference
     class OpenAiCompletions
-      def self.perform!(content, model = "gpt-3.5-turbo")
+      def self.perform!(messages, model = "gpt-3.5-turbo")
         headers = {
           "Authorization" => "Bearer #{SiteSetting.ai_openai_api_key}",
           "Content-Type" => "application/json",
@@ -14,7 +14,7 @@ module ::DiscourseAi
         response =
           Faraday.new(nil, connection_opts).post(
             "https://api.openai.com/v1/chat/completions",
-            { model: model, messages: content }.to_json,
+            { model: model, messages: messages }.to_json,
             headers,
           )
 

spec/models/completion_prompt_spec.rb

@@ -0,0 +1,29 @@
+# frozen_string_literal: true
+
+RSpec.describe CompletionPrompt do
+  describe "validations" do
+    context "when there are too many messages" do
+      it "doesn't accept more than 20 messages" do
+        prompt = described_class.new(messages: [{ role: "system", content: "a" }] * 21)
+
+        expect(prompt.valid?).to eq(false)
+      end
+    end
+
+    context "when the message is over the max length" do
+      it "doesn't accept messages when the length is more than 1000 characters" do
+        prompt = described_class.new(messages: [{ role: "system", content: "a" * 1001 }])
+
+        expect(prompt.valid?).to eq(false)
+      end
+    end
+
+    context "when the message has invalid roles" do
+      it "doesn't accept messages when the role is invalid" do
+        prompt = described_class.new(messages: [{ role: "invalid", content: "a" }])
+
+        expect(prompt.valid?).to eq(false)
+      end
+    end
+  end
+end

spec/support/openai_completions_inference_stubs.rb

@@ -83,12 +83,11 @@ class OpenAiCompletionsInferenceStubs
       text =
         type == DiscourseAi::AiHelper::OpenAiPrompt::TRANSLATE ? spanish_text : translated_response
 
-      used_prompt = CompletionPrompt.find_by(name: type)
-      prompt = [{ role: "system", content: used_prompt.value }, { role: "user", content: text }]
+      prompt_messages = CompletionPrompt.find_by(name: type).messages_with_user_input(text)
 
       WebMock
         .stub_request(:post, "https://api.openai.com/v1/chat/completions")
-        .with(body: JSON.dump(model: "gpt-3.5-turbo", messages: prompt))
+        .with(body: { model: "gpt-3.5-turbo", messages: prompt_messages }.to_json)
         .to_return(status: 200, body: JSON.dump(response(response_text_for(type))))
     end
   end