FEATURE: Use Persona's when scanning posts for spam

Author: romanrizzi

app/controllers/discourse_ai/admin/ai_spam_controller.rb

@@ -11,6 +11,13 @@ def show
 
       def update
         initial_settings = AiModerationSetting.spam
+
+        initial_data = {
+          custom_instructions: initial_settings&.data&.dig("custom_instructions"),
+          llm_model_id: initial_settings&.llm_model_id,
+          ai_persona_id: initial_settings&.ai_persona_id,
+        }
+
         initial_custom_instructions = initial_settings&.data&.dig("custom_instructions")
         initial_llm_model_id = initial_settings&.llm_model_id
 
@@ -29,6 +36,22 @@ def update
             )
           end
         end
+
+        if allowed_params.key?(:ai_persona_id)
+          updated_params[:ai_persona_id] = allowed_params[:ai_persona_id]
+          persona = AiPersona.find_by(id: allowed_params[:ai_persona_id])
+          if persona.nil? ||
+               persona.response_format.to_a.none? { |rf|
+                 rf["key"] == "spam" && rf["type"] == "boolean"
+               }
+            return(
+              render_json_error(
+                I18n.t("discourse_ai.llm.configuration.invalid_persona_response_format"),
+                status: 422,
+              )
+            )
+          end
+        end
         updated_params[:data] = {
           custom_instructions: allowed_params[:custom_instructions],
         } if allowed_params.key?(:custom_instructions)
@@ -41,7 +64,7 @@ def update
             AiModerationSetting.create!(updated_params.merge(setting_type: :spam))
           end
 
-          log_ai_spam_update(initial_llm_model_id, initial_custom_instructions, allowed_params)
+          log_ai_spam_update(initial_data, allowed_params)
         end
 
         is_enabled = ActiveModel::Type::Boolean.new.cast(allowed_params[:is_enabled])
@@ -119,9 +142,10 @@ def fix_errors
 
       private
 
-      def log_ai_spam_update(initial_llm_model_id, initial_custom_instructions, params)
+      def log_ai_spam_update(initial_data, params)
         changes_to_log = {}
 
+        initial_llm_model_id = initial_data[:llm_model_id]
         if params.key?(:llm_model_id) && initial_llm_model_id.to_s != params[:llm_model_id].to_s
           old_model_name =
             LlmModel.find_by(id: initial_llm_model_id)&.display_name || initial_llm_model_id
@@ -131,11 +155,22 @@ def log_ai_spam_update(initial_llm_model_id, initial_custom_instructions, params
           changes_to_log[:llm_model_id] = "#{old_model_name} → #{new_model_name}"
         end
 
+        initial_custom_instructions = initial_data[:custom_instructions]
         if params.key?(:custom_instructions) &&
              initial_custom_instructions != params[:custom_instructions]
           changes_to_log[:custom_instructions] = params[:custom_instructions]
         end
 
+        initial_ai_persona_id = initial_data[:ai_persona_id]
+        if params.key?(:ai_persona_id) && initial_ai_persona_id.to_s != params[:ai_persona_id].to_s
+          old_persona_name =
+            AiPersona.find_by(id: initial_ai_persona_id)&.name || initial_ai_persona_id
+          new_persona_name =
+            AiPersona.find_by(id: params[:ai_persona_id])&.name || params[:ai_persona_id]
+
+          changes_to_log[:ai_persona_id] = "#{old_persona_name} → #{new_persona_name}"
+        end
+
         if changes_to_log.present?
           changes_to_log[:subject] = I18n.t("discourse_ai.spam_detection.logging_subject")
           logger = DiscourseAi::Utils::AiStaffActionLogger.new(current_user)
@@ -144,7 +179,7 @@ def log_ai_spam_update(initial_llm_model_id, initial_custom_instructions, params
       end
 
       def allowed_params
-        params.permit(:is_enabled, :llm_model_id, :custom_instructions)
+        params.permit(:is_enabled, :llm_model_id, :custom_instructions, :ai_persona_id)
       end
 
       def spam_config

app/models/ai_moderation_setting.rb

@@ -1,6 +1,7 @@
 # frozen_string_literal: true
 class AiModerationSetting < ActiveRecord::Base
   belongs_to :llm_model
+  belongs_to :ai_persona
 
   validates :llm_model_id, presence: true
   validates :setting_type, presence: true
@@ -19,12 +20,13 @@ def custom_instructions
 #
 # Table name: ai_moderation_settings
 #
-#  id           :bigint           not null, primary key
-#  setting_type :enum             not null
-#  data         :jsonb
-#  llm_model_id :bigint           not null
-#  created_at   :datetime         not null
-#  updated_at   :datetime         not null
+#  id            :bigint           not null, primary key
+#  setting_type  :enum             not null
+#  data          :jsonb
+#  llm_model_id  :bigint           not null
+#  created_at    :datetime         not null
+#  updated_at    :datetime         not null
+#  ai_persona_id :bigint           default(-31), not null
 #
 # Indexes
 #

app/serializers/ai_spam_serializer.rb

@@ -8,7 +8,9 @@ class AiSpamSerializer < ApplicationSerializer
              :stats,
              :flagging_username,
              :spam_score_type,
-             :spam_scanning_user
+             :spam_scanning_user,
+             :ai_persona_id,
+             :available_personas
 
   def is_enabled
     object[:enabled]
@@ -18,6 +20,11 @@ def llm_id
     settings&.llm_model&.id
   end
 
+  def ai_persona_id
+    settings&.ai_persona&.id ||
+      DiscourseAi::Personas::Persona.system_personas[DiscourseAi::Personas::SpamDetector]
+  end
+
   def custom_instructions
     settings&.custom_instructions
   end
@@ -28,6 +35,12 @@ def available_llms
       .map { |hash| { id: hash[:value], name: hash[:name] } }
   end
 
+  def available_personas
+    DiscourseAi::Configuration::PersonaEnumerator.values.map do |h|
+      { id: h[:value], name: h[:name] }
+    end
+  end
+
   def flagging_username
     object[:flagging_username]
   end

assets/javascripts/discourse/components/ai-spam.gjs

@@ -35,6 +35,7 @@ export default class AiSpam extends Component {
   };
   @tracked isEnabled = false;
   @tracked selectedLLM = null;
+  @tracked selectedPersonaId = null;
   @tracked customInstructions = "";
   @tracked errors = [];
 
@@ -98,6 +99,7 @@ export default class AiSpam extends Component {
     }
     this.customInstructions = model.custom_instructions;
     this.stats = model.stats;
+    this.selectedPersonaId = model.ai_persona_id;
   }
 
   get availableLLMs() {
@@ -133,6 +135,11 @@ export default class AiSpam extends Component {
     this.selectedLLM = value;
   }
 
+  @action
+  async updatePersona(value) {
+    this.selectedPersonaId = value;
+  }
+
   @action
   async save() {
     try {
@@ -141,6 +148,7 @@ export default class AiSpam extends Component {
         data: {
           llm_model_id: this.llmId,
           custom_instructions: this.customInstructions,
+          ai_persona_id: this.selectedPersonaId,
         },
       });
       this.toasts.success({
@@ -256,6 +264,18 @@ export default class AiSpam extends Component {
           {{/if}}
         </div>
 
+        <div class="ai-spam__persona">
+          <label class="ai-spam__persona-label">{{i18n
+              "discourse_ai.spam.select_persona"
+            }}</label>
+          <ComboBox
+            @value={{this.selectedPersonaId}}
+            @content={{@model.available_personas}}
+            @onChange={{this.updatePersona}}
+            class="ai-spam__persona-selector"
+          />
+        </div>
+
         <div class="ai-spam__instructions">
           <label class="ai-spam__instructions-label">
             {{i18n "discourse_ai.spam.custom_instructions"}}

assets/stylesheets/modules/llms/common/spam.scss

@@ -24,12 +24,14 @@
 
   &__toggle,
   &__llm,
+  &__persona,
   &__instructions {
     margin-bottom: 1em;
   }
 
   &__toggle-label,
   &__llm-label,
+  &__persona-label,
   &__instructions-label {
     display: block;
     margin-bottom: 0.5em;

config/locales/client.en.yml

@@ -247,6 +247,7 @@ en:
         short_title: "Spam"
         title: "Configure spam handling"
         select_llm: "Select LLM"
+        select_persona: "Select persona"
         custom_instructions: "Custom instructions"
         custom_instructions_help: "Custom instructions specific to your site to help guide the AI in identifying spam, e.g. 'Be more aggressive about scanning posts not in English'."
         last_seven_days: "Last 7 days"

config/locales/server.en.yml

@@ -391,6 +391,9 @@ en:
         short_text_translator:
           name: "Short text translator"
           description: "Powers the translation feature by as a generic text translator, used for short texts like category names or tags"
+        spam_detector:
+          name: "Spam detector"
+          description: "Default persona powering our Spam detection feature"
 
       topic_not_found: "Summary unavailable, topic not found!"
       summarizing: "Summarizing topic"
@@ -577,6 +580,7 @@ en:
         set_llm_first: "Set %{setting} first"
         model_unreachable: "We couldn't get a response from this model. Check your settings first."
         invalid_seeded_model: "You can't use this model with this feature"
+        invalid_persona_response_format: "The selected persona must have a response format with a boolean field names \"spam\""
         must_select_model: "You must select a LLM first"
       endpoints:
         not_configured: "%{display_name} (not configured)"

db/migrate/20250619105705_add_persona_to_ai_moderation_settings.rb

@@ -0,0 +1,6 @@
+# frozen_string_literal: true
+class AddPersonaToAiModerationSettings < ActiveRecord::Migration[7.2]
+  def change
+    add_column :ai_moderation_settings, :ai_persona_id, :bigint, null: false, default: -31
+  end
+end