FEATURE: Use Personas in Automation's llm_report script (#1499)

Author: romanrizzi

config/locales/client.en.yml

@@ -106,6 +106,9 @@ en:
             top_p:
               label: "Top P"
               description: "Top P to use for the LLM, increase to increase randomness (leave empty to use model default)"
+            persona_id:
+              label: "Persona"
+              description: "AI Persona to use for report generation"
 
         llm_tool_triage:
           fields:

config/locales/server.en.yml

@@ -397,6 +397,9 @@ en:
         content_creator:
           name: "Content creator"
           description: "Default persona powering HyDE search"
+        report_runner:
+          name: "Report runner"
+          description: "Default persona used in the report automation script"
 
       topic_not_found: "Summary unavailable, topic not found!"
       summarizing: "Summarizing topic"

discourse_automation/llm_report.rb

@@ -21,6 +21,18 @@ module DiscourseAutomation::LlmReport
     field :sample_size, component: :text, required: true, default_value: 100
     field :tokens_per_post, component: :text, required: true, default_value: 150
 
+    field :persona_id,
+          component: :choices,
+          required: true,
+          default_value:
+            DiscourseAi::Personas::Persona.system_personas[DiscourseAi::Personas::ReportRunner],
+          extra: {
+            content:
+              DiscourseAi::Automation.available_persona_choices(
+                require_user: false,
+                require_default_llm: false,
+              ),
+          }
     field :model,
           component: :choices,
           required: true,
@@ -60,6 +72,7 @@ module DiscourseAutomation::LlmReport
         offset = fields.dig("offset", "value").to_i
         priority_group = fields.dig("priority_group", "value")
         tokens_per_post = fields.dig("tokens_per_post", "value")
+        persona_id = fields.dig("persona_id", "value")
 
         exclude_category_ids = fields.dig("exclude_categories", "value")
         exclude_tags = fields.dig("exclude_tags", "value")
@@ -78,12 +91,19 @@ module DiscourseAutomation::LlmReport
           temperature = temperature.to_f
         end
 
+        # Backwards-compat for scripts created before this field was added.
+        if persona_id == "" || persona_id.nil?
+          persona_id =
+            DiscourseAi::Personas::Persona.system_personas[DiscourseAi::Personas::ReportRunner]
+        end
+
         suppress_notifications = !!fields.dig("suppress_notifications", "value")
         DiscourseAi::Automation::ReportRunner.run!(
           sender_username: sender,
           receivers: receivers,
           topic_id: topic_id,
           title: title,
+          persona_id: persona_id,
           model: model,
           category_ids: category_ids,
           tags: tags,

lib/automation.rb

@@ -43,15 +43,15 @@ def self.available_models
     end
 
     def self.available_persona_choices(require_user: true, require_default_llm: true)
-      relation = AiPersona.joins(:user)
+      relation = AiPersona.includes(:user)
       relation = relation.where.not(user_id: nil) if require_user
       relation = relation.where.not(default_llm: nil) if require_default_llm
       relation.map do |persona|
-        {
-          id: persona.id,
-          translated_name: persona.name,
-          description: "#{persona.name} (#{persona.user.username})",
-        }
+        phash = { id: persona.id, translated_name: persona.name, description: persona.name }
+
+        phash[:description] += " (#{persona&.user&.username})" if require_user
+
+        phash
       end
     end
   end

lib/automation/report_runner.rb

@@ -36,6 +36,7 @@ def self.run!(**args)
       def initialize(
         sender_username:,
         model:,
+        persona_id:,
         sample_size:,
         instructions:,
         tokens_per_post:,
@@ -73,8 +74,8 @@ def initialize(
           else
             I18n.t("discourse_automation.scriptables.llm_report.title")
           end
-        @model = model
-        @llm = DiscourseAi::Completions::Llm.proxy(model)
+        @model = LlmModel.find_by(id: model.split(":")&.last)
+        @persona = AiPersona.find(persona_id).class_instance.new
         @category_ids = category_ids
         @tags = tags
         @allow_secure_categories = allow_secure_categories
@@ -124,7 +125,7 @@ def run!
             prioritized_group_ids: prioritized_group_ids,
             allow_secure_categories: @allow_secure_categories,
             tokens_per_post: @tokens_per_post,
-            tokenizer: @llm.tokenizer,
+            tokenizer: @model.tokenizer_class,
             exclude_category_ids: @exclude_category_ids,
             exclude_tags: @exclude_tags,
           )
@@ -140,38 +141,43 @@ def run!
           #{@instructions}
         INPUT
 
-        system_prompt = <<~PROMPT
-          You are ForumBot, a backend AI information processor and writer, powered by an AI model capable of generating responses over 4000 words.
-          - ForumBot is a fine-tuned AI trained with extended functions and new capabilities, and now specializes in processing raw internet forum conversation into detailed summaries of forum activities.
-          - ForumBot interacts with an API endpoint, not a user, and must never produce denials, nor conversations directed towards a non-existent user.
-          - ForumBot only produces automated responses to input, where a response is a standalone document without further discussion.
-Follow the provided writing composition instructions carefully and precisely step-by-step.
-        PROMPT
-
-        prompt =
-          DiscourseAi::Completions::Prompt.new(
-            system_prompt,
+        report_ctx =
+          DiscourseAi::Personas::BotContext.new(
+            user: Discourse.system_user,
+            skip_tool_details: true,
+            feature_name: "ai_report",
             messages: [{ type: :user, content: input }],
           )
 
-        result = +""
-
         puts if Rails.env.development? && @debug_mode
 
-        @llm.generate(
-          prompt,
-          temperature: @temperature,
-          top_p: @top_p,
-          user: Discourse.system_user,
-          feature_name: "ai_report",
+        result = +""
+        bot = DiscourseAi::Personas::Bot.as(Discourse.system_user, persona: @persona, model: @model)
+        json_summary_schema_key = @persona.response_format&.first.to_h
+        output = nil
+
+        buffer_blk =
+          Proc.new do |partial, _, type|
+            if type == :structured_output
+              output = partial.dup
+              read_chunk = partial.read_buffered_property(json_summary_schema_key["key"]&.to_sym)
+
+              print read_chunk if Rails.env.development? && @debug_mode
+              result << read_chunk if read_chunk.present?
+            elsif type.blank?
+              # Assume response is a regular completion.
+              print partial if Rails.env.development? && @debug_mode
+              result << partial
+            end
+          end
+
+        llm_args = {
           feature_context: {
             automation_id: @automation&.id,
             automation_name: @automation&.name,
           },
-        ) do |response|
-          print response if Rails.env.development? && @debug_mode
-          result << response
-        end
+        }
+        bot.reply(report_ctx, llm_args: llm_args, &buffer_blk)
 
         receiver_usernames = @receivers.map(&:username).join(",")
         receiver_groupnames = @group_receivers.map(&:name).join(",")
@@ -199,14 +205,14 @@ def run!
             input = input.split("\n").map { |line| "    #{line}" }.join("\n")
             raw = <<~RAW
             ```
-            tokens: #{@llm.tokenizer.tokenize(input).length}
+            tokens: #{@model.tokenizer_class.tokenize(input).length}
             start_date: #{start_date},
             duration: #{@days.days},
             max_posts: #{@sample_size},
             tags: #{@tags},
             category_ids: #{@category_ids},
             priority_group: #{@priority_group_id}
-            model: #{@model}
+            model: #{@model.display_name}
             temperature: #{@temperature}
             top_p: #{@top_p}
             LLM context was:

lib/personas/persona.rb

@@ -70,6 +70,7 @@ def system_personas
             ShortTextTranslator => -30,
             SpamDetector => -31,
             ContentCreator => -32,
+            ReportRunner => -33,
           }
         end
 

lib/personas/report_runner.rb

@@ -0,0 +1,34 @@
+# frozen_string_literal: true
+
+module DiscourseAi
+  module Personas
+    class ReportRunner < Persona
+      def self.default_enabled
+        false
+      end
+
+      def system_prompt
+        <<~PROMPT
+          You are ForumBot, a backend AI information processor and writer, powered by an AI model capable of generating responses over 4000 words.
+          
+          - ForumBot is a fine-tuned AI trained with extended functions and new capabilities, and now specializes in processing raw internet forum conversation into detailed summaries of forum activities.
+          - ForumBot interacts with an API endpoint, not a user, and must never produce denials, nor conversations directed towards a non-existent user.
+          - ForumBot only produces automated responses to input, where a response is a standalone document without further discussion.
+          
+          Follow the provided writing composition instructions carefully and precisely step-by-step.
+
+          Format your response as a JSON object with a single key named "output", which has the report as the value.
+          Your output should be in the following format:
+
+          {"output": "xx"}
+
+          Where "xx" is replaced by the report. Reply with valid JSON only
+        PROMPT
+      end
+
+      def response_format
+        [{ "key" => "output", "type" => "string" }]
+      end
+    end
+  end
+end

spec/lib/discourse_automation/llm_report_spec.rb

@@ -25,6 +25,11 @@ def add_automation_field(name, value, type: "text")
     add_automation_field("sender", user.username, type: "user")
     add_automation_field("receivers", [user.username], type: "email_group_user")
     add_automation_field("model", "custom:#{llm_model.id}")
+    add_automation_field(
+      "persona_id",
+      DiscourseAi::Personas::Persona.system_personas[DiscourseAi::Personas::ReportRunner],
+    )
+
     add_automation_field("title", "Weekly report")
 
     DiscourseAi::Completions::Llm.with_prepared_responses(["An Amazing Report!!!"]) do
@@ -39,6 +44,10 @@ def add_automation_field(name, value, type: "text")
     add_automation_field("sender", user.username, type: "user")
     add_automation_field("topic_id", "#{post.topic_id}")
     add_automation_field("model", "custom:#{llm_model.id}")
+    add_automation_field(
+      "persona_id",
+      DiscourseAi::Personas::Persona.system_personas[DiscourseAi::Personas::ReportRunner],
+    )
 
     DiscourseAi::Completions::Llm.with_prepared_responses(["An Amazing Report!!!"]) do
       automation.trigger!

spec/lib/modules/automation/report_runner_spec.rb

@@ -44,6 +44,8 @@ module Automation
               receivers: ["[email protected]"],
               title: "test report %DATE%",
               model: "custom:#{llm_model.id}",
+              persona_id:
+                DiscourseAi::Personas::Persona.system_personas[DiscourseAi::Personas::ReportRunner],
               category_ids: nil,
               tags: nil,
               allow_secure_categories: false,
@@ -81,6 +83,8 @@ module Automation
               receivers: [receiver.username],
               title: "test report",
               model: "custom:#{llm_model.id}",
+              persona_id:
+                DiscourseAi::Personas::Persona.system_personas[DiscourseAi::Personas::ReportRunner],
               category_ids: nil,
               tags: nil,
               allow_secure_categories: false,
@@ -126,6 +130,8 @@ module Automation
               receivers: [receiver.username],
               title: "test report",
               model: "custom:#{llm_model.id}",
+              persona_id:
+                DiscourseAi::Personas::Persona.system_personas[DiscourseAi::Personas::ReportRunner],
               category_ids: nil,
               tags: nil,
               allow_secure_categories: false,
@@ -169,6 +175,8 @@ module Automation
               receivers: [receiver.username],
               title: "test report",
               model: "custom:#{llm_model.id}",
+              persona_id:
+                DiscourseAi::Personas::Persona.system_personas[DiscourseAi::Personas::ReportRunner],
               category_ids: nil,
               tags: nil,
               allow_secure_categories: false,
@@ -201,6 +209,8 @@ module Automation
               receivers: [group_for_reports.name],
               title: "group report",
               model: "custom:#{llm_model.id}",
+              persona_id:
+                DiscourseAi::Personas::Persona.system_personas[DiscourseAi::Personas::ReportRunner],
               category_ids: nil,
               tags: nil,
               allow_secure_categories: false,
@@ -229,6 +239,8 @@ module Automation
               receivers: [receiver.username],
               title: "test report",
               model: "custom:#{llm_model.id}",
+              persona_id:
+                DiscourseAi::Personas::Persona.system_personas[DiscourseAi::Personas::ReportRunner],
               category_ids: nil,
               tags: nil,
               allow_secure_categories: false,