discourse · romanrizzi · Aug 4, 2025 · Jul 31, 2025 · Jul 31, 2025
diff --git a/plugins/discourse-ai/config/locales/client.en.yml b/plugins/discourse-ai/config/locales/client.en.yml
@@ -132,9 +132,9 @@ en:
               description: "In silent mode persona will receive the content but will not post anything on the forum - useful when performing triage using tools"
         llm_triage:
           fields:
-            system_prompt:
-              label: "System Prompt"
-              description: "The prompt that will be used to triage, be sure for it to reply with a single word you can use to trigger the action"
+            triage_persona:
+              label: "Persona"
+              description: "Persona used to triage, be sure for it to reply with a single word you can use to trigger the action"
             max_post_tokens:
               label: "Max Post Tokens"
               description: "The maximum number of tokens to scan using LLM triage"
@@ -174,12 +174,6 @@ en:
             reply_persona:
               label: "Reply Persona"
               description: "AI Persona to use for replies (must have default LLM), will be prioritized over canned reply"
-            model:
-              label: "Model"
-              description: "Language model used for triage"
-            temperature:
-              label: "Temperature"
-              description: "Temperature to use for the LLM. Increase to increase randomness (leave empty to use model default)"
             max_output_tokens:
               label: "Max output tokens"
               description: "When specified, sets an upper bound to the maximum number of tokens the model can generate. Respects LLM's max output tokens limit"

diff --git a/plugins/discourse-ai/db/migrate/20250721080444_seed_personas_from_triage_scripts.rb b/plugins/discourse-ai/db/migrate/20250721080444_seed_personas_from_triage_scripts.rb
@@ -0,0 +1,70 @@
+# frozen_string_literal: true
+
+class SeedPersonasFromTriageScripts < ActiveRecord::Migration[8.0]
+  def up
+    script_fields = DB.query <<~SQL
+      SELECT fields.id, fields.name, (fields.metadata->>'value') AS value, automations.name AS automation_name, automations.id AS automation_id
+      FROM discourse_automation_fields fields
+      INNER JOIN discourse_automation_automations automations ON automations.id = fields.automation_id
+      WHERE fields.name IN ('model', 'system_prompt', 'temperature')
+      AND automations.script = 'llm_triage'
+    SQL
+
+    return if script_fields.empty?
+
+    script_fields =
+      script_fields.reduce({}) do |acc, field|
+        id = field.automation_id
+        acc[id] = { "automation_id" => id, "name" => field.automation_name } if acc[id].nil?
+
+        acc[field.automation_id].merge!(field.name => field.value)
+
+        acc
+      end
+
+    automation_to_persona_ids =
+      script_fields.transform_values do |field|
+        if field["system_prompt"].blank?
+          nil
+        else
+          name =
+            (
+              if field["name"].present?
+                "#{field["name"]} triage automation"
+              else
+                "Unnamed triage automation script ID #{field["automation_id"]}"
+              end
+            )
+          temp = field["temperature"] || "NULL"
+          row =
+            "'#{name}', 'Seeded Persona for an LLM Triage script', FALSE, '#{field["system_prompt"]}', #{temp}, #{field["model"]}, NOW(), NOW()"
+
+          DB.query_single(<<~SQL)&.first
+            INSERT INTO ai_personas (name, description, enabled, system_prompt, temperature, default_llm_id, created_at, updated_at)
+            VALUES (#{row})
+            RETURNING id
+          SQL
+        end
+      end
+
+    new_fields =
+      automation_to_persona_ids
+        .map do |k, v|
+          if v.blank?
+            nil
+          else
+            "(#{k}, 'triage_persona', json_build_object('value', #{v}), 'choices', 'script', NOW(), NOW())"
+          end
+        end
+        .compact
+
+    DB.exec <<~SQL
+      INSERT INTO discourse_automation_fields (automation_id, name, metadata, component, target, created_at, updated_at)
+      VALUES #{new_fields.join(",")}
+    SQL
+  end
+
+  def down
+    raise ActiveRecord::IrreversibleMigration
+  end
+end
diff --git a/plugins/discourse-ai/discourse_automation/llm_triage.rb b/plugins/discourse-ai/discourse_automation/llm_triage.rb
@@ -13,17 +13,19 @@
     field :include_personal_messages, component: :boolean
 
     # Inputs
-    field :model,
+    field :triage_persona,
           component: :choices,
           required: true,
           extra: {
-            content: DiscourseAi::Automation.available_models,
+            content:
+              DiscourseAi::Automation.available_persona_choices(
+                require_user: false,
+                require_default_llm: true,
+              ),
           }
-    field :system_prompt, component: :message, required: false
     field :search_for_text, component: :text, required: true
     field :max_post_tokens, component: :text
     field :stop_sequences, component: :text_list, required: false
-    field :temperature, component: :text
     field :max_output_tokens, component: :text
 
     # Actions
@@ -60,6 +62,8 @@
         next if !include_personal_messages
       end
 
+      triage_persona_id = fields.dig("triage_persona", "value")
+
       canned_reply = fields.dig("canned_reply", "value")
       canned_reply_user = fields.dig("canned_reply_user", "value")
       reply_persona_id = fields.dig("reply_persona", "value")
@@ -69,22 +73,14 @@
       next if post.user.username == canned_reply_user
       next if post.raw.strip == canned_reply.to_s.strip
 
-      system_prompt = fields.dig("system_prompt", "value")
       search_for_text = fields.dig("search_for_text", "value")
-      model = fields.dig("model", "value")
 
       category_id = fields.dig("category", "value")
       tags = fields.dig("tags", "value")
       hide_topic = fields.dig("hide_topic", "value")
       flag_post = fields.dig("flag_post", "value")
       flag_type = fields.dig("flag_type", "value")
       max_post_tokens = fields.dig("max_post_tokens", "value").to_i
-      temperature = fields.dig("temperature", "value")
-      if temperature == "" || temperature.nil?
-        temperature = nil
-      else
-        temperature = temperature.to_f
-      end
 
       max_output_tokens = fields.dig("max_output_tokens", "value").to_i
       max_output_tokens = nil if max_output_tokens <= 0
@@ -110,9 +106,8 @@
 
         DiscourseAi::Automation::LlmTriage.handle(
           post: post,
-          model: model,
+          triage_persona_id: triage_persona_id,
           search_for_text: search_for_text,
-          system_prompt: system_prompt,
           category_id: category_id,
           tags: tags,
           canned_reply: canned_reply,
@@ -125,7 +120,6 @@
           max_post_tokens: max_post_tokens,
           stop_sequences: stop_sequences,
           automation: self.automation,
-          temperature: temperature,
           max_output_tokens: max_output_tokens,
           action: context["action"],
         )

diff --git a/plugins/discourse-ai/lib/ai_helper/assistant.rb b/plugins/discourse-ai/lib/ai_helper/assistant.rb
@@ -312,7 +312,7 @@ def find_ai_helper_model(helper_mode, persona_klass)
 
       # Priorities are:
       #   1. Persona's default LLM
-      #   2. SiteSetting.ai_default_llm_id (or newest LLM if not set)
+      #   2. SiteSetting.ai_default_llm_model (or newest LLM if not set)
       def self.find_ai_helper_model(helper_mode, persona_klass)
         model_id = persona_klass.default_llm_id || SiteSetting.ai_default_llm_model
 

diff --git a/plugins/discourse-ai/lib/automation/llm_triage.rb b/plugins/discourse-ai/lib/automation/llm_triage.rb
@@ -5,9 +5,8 @@ module Automation
     module LlmTriage
       def self.handle(
         post:,
-        model:,
+        triage_persona_id:,
         search_for_text:,
-        system_prompt:,
         category_id: nil,
         tags: nil,
         canned_reply: nil,
@@ -18,7 +17,6 @@ def self.handle(
         automation: nil,
         max_post_tokens: nil,
         stop_sequences: nil,
-        temperature: nil,
         whisper: nil,
         reply_persona_id: nil,
         max_output_tokens: nil,
@@ -34,42 +32,55 @@ def self.handle(
           return
         end
 
-        llm = DiscourseAi::Completions::Llm.proxy(model)
+        triage_persona = AiPersona.find(triage_persona_id)
+        model_id = triage_persona.default_llm_id || SiteSetting.ai_default_llm_model
+        return if model_id.blank?
+        model = LlmModel.find(model_id)
 
-        s_prompt = system_prompt.to_s.sub("%%POST%%", "") # Backwards-compat. We no longer sub this.
-        prompt = DiscourseAi::Completions::Prompt.new(s_prompt)
+        bot =
+          DiscourseAi::Personas::Bot.as(
+            Discourse.system_user,
+            persona: triage_persona.class_instance.new,
+            model: model,
+          )
 
-        content = "title: #{post.topic.title}\n#{post.raw}"
+        input = "title: #{post.topic.title}\n#{post.raw}"
 
-        content =
-          llm.tokenizer.truncate(
-            content,
+        input =
+          model.tokenizer_class.truncate(
+            input,
             max_post_tokens,
             strict: SiteSetting.ai_strict_token_counting,
           ) if max_post_tokens.present?
 
         if post.upload_ids.present?
-          content = [content]
-          content.concat(post.upload_ids.map { |upload_id| { upload_id: upload_id } })
+          input = [input]
+          input.concat(post.upload_ids.map { |upload_id| { upload_id: upload_id } })
         end
 
-        prompt.push(type: :user, content: content)
+        bot_ctx =
+          DiscourseAi::Personas::BotContext.new(
+            user: Discourse.system_user,
+            skip_tool_details: true,
+            feature_name: "llm_triage",
+            messages: [{ type: :user, content: input }],
+          )
 
         result = nil
 
-        result =
-          llm.generate(
-            prompt,
-            max_tokens: max_output_tokens,
-            temperature: temperature,
-            user: Discourse.system_user,
-            stop_sequences: stop_sequences,
-            feature_name: "llm_triage",
-            feature_context: {
-              automation_id: automation&.id,
-              automation_name: automation&.name,
-            },
-          )&.strip
+        llm_args = {
+          max_tokens: max_output_tokens,
+          stop_sequences: stop_sequences,
+          feature_context: {
+            automation_id: automation&.id,
+            automation_name: automation&.name,
+          },
+        }
+
+        result = +""
+        bot.reply(bot_ctx, llm_args: llm_args) do |partial, _, type|
+          result << partial if type.blank?
+        end
 
         if result.present? && result.downcase.include?(search_for_text.downcase)
           user = User.find_by_username(canned_reply_user) if canned_reply_user.present?
@@ -92,6 +103,7 @@ def self.handle(
             end
           elsif canned_reply.present? && action != :edit
             post_type = whisper ? Post.types[:whisper] : Post.types[:regular]
+
             PostCreator.create!(
               user,
               topic_id: post.topic_id,

diff --git a/plugins/discourse-ai/lib/embeddings/semantic_search.rb b/plugins/discourse-ai/lib/embeddings/semantic_search.rb
@@ -201,7 +201,7 @@ def hypothetical_post_from(search_term)
 
       # Priorities are:
       #   1. Persona's default LLM
-      #   2. SiteSetting.ai_default_llm_id (or newest LLM if not set)
+      #   2. SiteSetting.ai_default_llm_model (or newest LLM if not set)
       def find_ai_hyde_model(persona_klass)
         model_id = persona_klass.default_llm_id || SiteSetting.ai_default_llm_model
 

diff --git a/plugins/discourse-ai/lib/summarization.rb b/plugins/discourse-ai/lib/summarization.rb
@@ -54,7 +54,7 @@ def chat_channel_summary(channel, time_window_in_hours)
 
       # Priorities are:
       #   1. Persona's default LLM
-      #   2. SiteSetting.ai_default_llm_id (or newest LLM if not set)
+      #   2. SiteSetting.ai_default_llm_model (or newest LLM if not set)
       def find_summarization_model(persona_klass)
         model_id = persona_klass.default_llm_id || SiteSetting.ai_default_llm_model
 

diff --git a/plugins/discourse-ai/spec/lib/discourse_automation/llm_triage_spec.rb b/plugins/discourse-ai/spec/lib/discourse_automation/llm_triage_spec.rb
@@ -11,6 +11,7 @@
   let(:automation) { Fabricate(:automation, script: "llm_triage", enabled: true) }
 
   fab!(:llm_model)
+  fab!(:ai_persona)
 
   def add_automation_field(name, value, type: "text")
     automation.fields.create!(
@@ -27,9 +28,11 @@ def add_automation_field(name, value, type: "text")
     enable_current_plugin
 
     SiteSetting.tagging_enabled = true
-    add_automation_field("system_prompt", "hello %%POST%%")
+
+    ai_persona.update!(default_llm: llm_model)
+
+    add_automation_field("triage_persona", ai_persona.id)
     add_automation_field("search_for_text", "bad")
-    add_automation_field("model", llm_model.id)
     add_automation_field("category", category.id, type: "category")
     add_automation_field("tags", %w[aaa bbb], type: "tags")
     add_automation_field("hide_topic", true, type: "boolean")
@@ -42,21 +45,17 @@ def add_automation_field(name, value, type: "text")
   it "can trigger via automation" do
     post = Fabricate(:post, raw: "hello " * 5000)
 
-    body = {
-      model: "gpt-3.5-turbo-0301",
-      usage: {
-        prompt_tokens: 337,
-        completion_tokens: 162,
-        total_tokens: 499,
-      },
-      choices: [
-        { message: { role: "assistant", content: "bad" }, finish_reason: "stop", index: 0 },
-      ],
-    }.to_json
+    chunks = <<~RESPONSE
+    data: {"id":"chatcmpl-B2VwlY6KzSDtHvg8pN1VAfRhhLFgn","object":"chat.completion.chunk","created":1739939159,"model": "gpt-3.5-turbo-0301","service_tier":"default","system_fingerprint":"fp_ef58bd3122","choices":[{"index":0,"delta":{"role":"assistant","content":"","refusal":null},"finish_reason":null}],"usage":null}
+
+    data: {"id":"chatcmpl-B2VwlY6KzSDtHvg8pN1VAfRhhLFgn","object":"chat.completion.chunk","created":1739939159,"model": "gpt-3.5-turbo-0301","service_tier":"default","system_fingerprint":"fp_ef58bd3122","choices":[{"index":0,"delta":{"content":"bad"},"finish_reason":null}],"usage":null}
+
+    data: [DONE]
+    RESPONSE
 
     WebMock.stub_request(:post, "https://api.openai.com/v1/chat/completions").to_return(
       status: 200,
-      body: body,
+      body: chunks,
     )
 
     automation.running_in_background!
@@ -96,7 +95,7 @@ def add_automation_field(name, value, type: "text")
     # PM
     reply_user.update!(admin: true)
     add_automation_field("include_personal_messages", true, type: :boolean)
-    add_automation_field("temperature", "0.2")
+    ai_persona.update!(temperature: 0.2)
     add_automation_field("max_output_tokens", "700")
     post = Fabricate(:post, topic: personal_message)