FEATURE: Personas powered summaries. (#1232)

* REFACTOR: Move personas into it's own module.

* WIP: Use personas for summarization

* Prioritize persona default LLM or fallback to newest one

* Simplify summarization strategy

* Keep ai_sumarization_model as a fallback

app/controllers/discourse_ai/ai_bot/bot_controller.rb

@@ -47,9 +47,9 @@ module DiscourseAi
 
       def discover
         ai_persona =
-          AiPersona.all_personas.find do |persona|
+          AiPersona
-            persona.id == SiteSetting.ai_bot_discover_persona.to_i
+            .all_personas(enabled_only: false)
-          end
+            .find { |persona| persona.id == SiteSetting.ai_bot_discover_persona.to_i }
 
         if ai_persona.nil? || !current_user.in_any_groups?(ai_persona.allowed_group_ids.to_a)
           raise Discourse::InvalidAccess.new

app/jobs/regular/stream_discover_reply.rb

@@ -9,9 +9,9 @@ module Jobs
       return if (query = args[:query]).blank?
 
       ai_persona_klass =
-        AiPersona.all_personas.find do |persona|
+        AiPersona
-          persona.id == SiteSetting.ai_bot_discover_persona.to_i
+          .all_personas(enabled_only: false)
-        end
+          .find { |persona| persona.id == SiteSetting.ai_bot_discover_persona.to_i }
 
       if ai_persona_klass.nil? || !user.in_any_groups?(ai_persona_klass.allowed_group_ids.to_a)
         return

app/models/ai_persona.rb

@@ -46,13 +46,18 @@ class AiPersona < ActiveRecord::Base
 
   scope :ordered, -> { order("priority DESC, lower(name) ASC") }
 
-  def self.all_personas
+  def self.all_personas(enabled_only: true)
     persona_cache[:value] ||= AiPersona
       .ordered
-      .where(enabled: true)
       .all
       .limit(MAX_PERSONAS_PER_SITE)
       .map(&:class_instance)
+
+    if enabled_only
+      persona_cache[:value].select { |p| p.enabled }
+    else
+      persona_cache[:value]
+    end
   end
 
   def self.persona_users(user: nil)
@@ -176,6 +181,7 @@ class AiPersona < ActiveRecord::Base
       description
       allowed_group_ids
       tool_details
+      enabled
     ]
 
     instance_attributes = {}

config/locales/server.en.yml

@@ -303,6 +303,12 @@ en:
         web_artifact_creator:
           name: "Web Artifact Creator"
           description: "AI Bot specialized in creating interactive web artifacts"
+        summarizer:
+          name: "Summarizer"
+          description: "Default persona used to power AI summaries"
+        short_summarizer:
+          name: "Summarizer (short form)"
+          description: "Default persona used to power AI short summaries for topic lists' items"
       topic_not_found: "Summary unavailable, topic not found!"
       summarizing: "Summarizing topic"
       searching: "Searching for: '%{query}'"
@@ -452,6 +458,7 @@ en:
 
     llm:
       configuration:
+        create_llm: "You need to setup an LLM before enabling this feature"
         disable_module_first: "You have to disable %{setting} first."
         set_llm_first: "Set %{setting} first"
         model_unreachable: "We couldn't get a response from this model. Check your settings first."

config/settings.yml

@@ -240,18 +240,30 @@ discourse_ai:
     type: enum
     enum: "DiscourseAi::Configuration::LlmEnumerator"
     validator: "DiscourseAi::Configuration::LlmValidator"
+    hidden: true
+  ai_summarization_persona:
+    default: "-11"
+    type: enum
+    enum: "DiscourseAi::Configuration::PersonaEnumerator"
+
   ai_pm_summarization_allowed_groups:
     type: group_list
     list_type: compact
     default: ""
-  ai_custom_summarization_allowed_groups:
+  ai_custom_summarization_allowed_groups: # Deprecated. TODO(roman): Remove 2025-09-01
     type: group_list
     list_type: compact
     default: "3|13" # 3: @staff, 13: @trust_level_3
+    hidden: true
   ai_summary_gists_enabled:
     default: false
     hidden: true
-  ai_summary_gists_allowed_groups:
+  ai_summary_gists_persona:
+    default: "-12"
+    type: enum
+    enum: "DiscourseAi::Configuration::PersonaEnumerator"
+    hidden: true
+  ai_summary_gists_allowed_groups: # Deprecated. TODO(roman): Remove 2025-09-01
     type: group_list
     list_type: compact
     default: "0" #everyone

db/fixtures/personas/603_ai_personas.rb

@@ -1,17 +1,39 @@
 # frozen_string_literal: true
 
+summarization_personas = [DiscourseAi::Personas::Summarizer, DiscourseAi::Personas::ShortSummarizer]
+
+def from_setting(setting_name)
+  DB.query_single(
+    "SELECT value FROM site_settings WHERE name = :setting_name",
+    setting_name: setting_name,
+  )
+end
+
 DiscourseAi::Personas::Persona.system_personas.each do |persona_class, id|
   persona = AiPersona.find_by(id: id)
   if !persona
     persona = AiPersona.new
     persona.id = id
+
     if persona_class == DiscourseAi::Personas::WebArtifactCreator
       # this is somewhat sensitive, so we default it to staff
       persona.allowed_group_ids = [Group::AUTO_GROUPS[:staff]]
+    elsif summarization_personas.include?(persona_class)
+      # Copy group permissions from site settings.
+      default_groups = [Group::AUTO_GROUPS[:staff], Group::AUTO_GROUPS[:trust_level_3]]
+
+      setting_name = "ai_custom_summarization_allowed_groups"
+      if persona_class == DiscourseAi::Personas::ShortSummarizer
+        setting_name = "ai_summary_gists_allowed_groups"
+        default_groups = [] # Blank == everyone
+      end
+
+      persona.allowed_group_ids = from_setting(setting_name).first&.split("|") || default_groups
     else
       persona.allowed_group_ids = [Group::AUTO_GROUPS[:trust_level_0]]
     end
-    persona.enabled = true
+
+    persona.enabled = !summarization_personas.include?(persona_class)
     persona.priority = true if persona_class == DiscourseAi::Personas::General
   end
 

lib/configuration/llm_dependency_validator.rb

@@ -10,17 +10,27 @@ module DiscourseAi
       def valid_value?(val)
         return true if val == "f"
 
+        if @opts[:name] == :ai_summarization_enabled
+          has_llms = LlmModel.count > 0
+          @no_llms_configured = !has_llms
+          has_llms
+        else
           @llm_dependency_setting_name =
             DiscourseAi::Configuration::LlmValidator.new.choose_llm_setting_for(@opts[:name])
 
           SiteSetting.public_send(@llm_dependency_setting_name).present?
         end
+      end
 
       def error_message
+        if @llm_dependency_setting_name
           I18n.t(
             "discourse_ai.llm.configuration.set_llm_first",
             setting: @llm_dependency_setting_name,
           )
+        elsif @no_llms_configured
+          I18n.t("discourse_ai.llm.configuration.create_llm")
+        end
       end
     end
   end

lib/configuration/llm_enumerator.rb

@@ -26,7 +26,9 @@ module DiscourseAi
         end
 
         if SiteSetting.ai_summarization_enabled
-          model_id = SiteSetting.ai_summarization_model.split(":").last.to_i
+          summarization_persona = AiPersona.find_by(id: SiteSetting.ai_summarization_persona)
+          model_id = summarization_persona.default_llm_id || LlmModel.last&.id
+
           rval[model_id] << { type: :ai_summarization }
         end
 

lib/configuration/persona_enumerator.rb

@@ -10,7 +10,9 @@ module DiscourseAi
       end
 
       def self.values
-        AiPersona.all_personas.map { |persona| { name: persona.name, value: persona.id } }
+        AiPersona
+          .all_personas(enabled_only: false)
+          .map { |persona| { name: persona.name, value: persona.id } }
       end
     end
   end

lib/discord/bot/persona_replier.rb

@@ -6,7 +6,7 @@ module DiscourseAi
       def initialize(body)
         @persona =
           AiPersona
-            .all_personas
+            .all_personas(enabled_only: false)
             .find { |persona| persona.id == SiteSetting.ai_discord_search_persona.to_i }
             .new
         @bot =

lib/guardian_extensions.rb

@@ -24,24 +24,26 @@ module DiscourseAi
     def can_see_gists?
       return false if !SiteSetting.ai_summarization_enabled
       return false if !SiteSetting.ai_summary_gists_enabled
-      if SiteSetting.ai_summary_gists_allowed_groups.to_s == Group::AUTO_GROUPS[:everyone].to_s
-        return true
-      end
-      return false if anonymous?
-      return false if SiteSetting.ai_summary_gists_allowed_groups_map.empty?
 
-      SiteSetting.ai_summary_gists_allowed_groups_map.any? do |group_id|
+      if (ai_persona = AiPersona.find_by(id: SiteSetting.ai_summary_gists_persona)).blank?
-        user.group_ids.include?(group_id)
+        return false
       end
+      persona_groups = ai_persona.allowed_group_ids.to_a
+      return true if persona_groups.empty?
+      return false if anonymous?
+
+      ai_persona.allowed_group_ids.to_a.any? { |group_id| user.group_ids.include?(group_id) }
     end
 
     def can_request_summary?
       return false if anonymous?
 
       user_group_ids = user.group_ids
-      SiteSetting.ai_custom_summarization_allowed_groups_map.any? do |group_id|
+      if (ai_persona = AiPersona.find_by(id: SiteSetting.ai_summarization_persona)).blank?
-        user_group_ids.include?(group_id)
+        return false
       end
+
+      ai_persona.allowed_group_ids.to_a.any? { |group_id| user.group_ids.include?(group_id) }
     end
 
     def can_debug_ai_bot_conversation?(target)

lib/personas/bot.rb

@@ -28,7 +28,7 @@ module DiscourseAi
       attr_accessor :persona
 
       def llm
-        @llm ||= DiscourseAi::Completions::Llm.proxy(model)
+        DiscourseAi::Completions::Llm.proxy(model)
       end
 
       def force_tool_if_needed(prompt, context)
@@ -51,12 +51,12 @@ module DiscourseAi
         end
       end
 
-      def reply(context, &update_blk)
+      def reply(context, llm_args: {}, &update_blk)
         unless context.is_a?(BotContext)
           raise ArgumentError, "context must be an instance of BotContext"
         end
-        llm = DiscourseAi::Completions::Llm.proxy(model)
+        current_llm = llm
-        prompt = persona.craft_prompt(context, llm: llm)
+        prompt = persona.craft_prompt(context, llm: current_llm)
 
         total_completions = 0
         ongoing_chain = true
@@ -67,6 +67,7 @@ module DiscourseAi
         llm_kwargs = { user: user }
         llm_kwargs[:temperature] = persona.temperature if persona.temperature
         llm_kwargs[:top_p] = persona.top_p if persona.top_p
+        llm_kwargs[:max_tokens] = llm_args[:max_tokens] if llm_args[:max_tokens].present?
 
         needs_newlines = false
         tools_ran = 0
@@ -82,9 +83,9 @@ module DiscourseAi
           current_thinking = []
 
           result =
-            llm.generate(
+            current_llm.generate(
               prompt,
-              feature_name: "bot",
+              feature_name: context.feature_name,
               partial_tool_calls: allow_partial_tool_calls,
               output_thinking: true,
               **llm_kwargs,
@@ -93,7 +94,7 @@ module DiscourseAi
                 persona.find_tool(
                   partial,
                   bot_user: user,
-                  llm: llm,
+                  llm: current_llm,
                   context: context,
                   existing_tools: existing_tools,
                 )
@@ -120,7 +121,7 @@ module DiscourseAi
                 process_tool(
                   tool: tool,
                   raw_context: raw_context,
-                  llm: llm,
+                  current_llm: current_llm,
                   cancel: cancel,
                   update_blk: update_blk,
                   prompt: prompt,
@@ -204,7 +205,7 @@ module DiscourseAi
       def process_tool(
         tool:,
         raw_context:,
-        llm:,
+        current_llm:,
         cancel:,
         update_blk:,
         prompt:,
@@ -212,7 +213,7 @@ module DiscourseAi
         current_thinking:
       )
         tool_call_id = tool.tool_call_id
-        invocation_result_json = invoke_tool(tool, llm, cancel, context, &update_blk).to_json
+        invocation_result_json = invoke_tool(tool, cancel, context, &update_blk).to_json
 
         tool_call_message = {
           type: :tool_call,
@@ -246,7 +247,7 @@ module DiscourseAi
         raw_context << [invocation_result_json, tool_call_id, "tool", tool.name]
       end
 
-      def invoke_tool(tool, llm, cancel, context, &update_blk)
+      def invoke_tool(tool, cancel, context, &update_blk)
         show_placeholder = !context.skip_tool_details && !tool.class.allow_partial_tool_calls?
 
         update_blk.call("", cancel, build_placeholder(tool.summary, "")) if show_placeholder

lib/personas/bot_context.rb

@@ -14,7 +14,9 @@ module DiscourseAi
                     :chosen_tools,
                     :message_id,
                     :channel_id,
-                    :context_post_ids
+                    :context_post_ids,
+                    :feature_name,
+                    :resource_url
 
       def initialize(
         post: nil,
@@ -29,7 +31,9 @@ module DiscourseAi
         time: nil,
         message_id: nil,
         channel_id: nil,
-        context_post_ids: nil
+        context_post_ids: nil,
+        feature_name: "bot",
+        resource_url: nil
       )
         @participants = participants
         @user = user
@@ -45,6 +49,7 @@ module DiscourseAi
         @site_title = site_title
         @site_description = site_description
         @time = time
+        @feature_name = feature_name
 
         if post
           @post_id = post.id
@@ -56,7 +61,7 @@ module DiscourseAi
       end
 
       # these are strings that can be safely interpolated into templates
-      TEMPLATE_PARAMS = %w[time site_url site_title site_description participants]
+      TEMPLATE_PARAMS = %w[time site_url site_title site_description participants resource_url]
 
       def lookup_template_param(key)
         public_send(key.to_sym) if TEMPLATE_PARAMS.include?(key)
@@ -100,6 +105,8 @@ module DiscourseAi
           site_title: @site_title,
           site_description: @site_description,
           skip_tool_details: @skip_tool_details,
+          feature_name: @feature_name,
+          resource_url: @resource_url,
         }
       end
     end

lib/personas/persona.rb

@@ -44,6 +44,8 @@ module DiscourseAi
             DiscourseHelper => -8,
             GithubHelper => -9,
             WebArtifactCreator => -10,
+            Summarizer => -11,
+            ShortSummarizer => -12,
           }
         end
 

lib/personas/short_summarizer.rb

@@ -0,0 +1,26 @@
+# frozen_string_literal: true
+
+module DiscourseAi
+  module Personas
+    class ShortSummarizer < Persona
+      def system_prompt
+        <<~PROMPT.strip
+          You are an advanced summarization bot. Analyze a given conversation and produce a concise,
+          single-sentence summary that conveys the main topic and current developments to someone with no prior context.
+
+          ### Guidelines:
+
+          - Emphasize the most recent updates while considering their significance within the original post.
+          - Focus on the central theme or issue being addressed, maintaining an objective and neutral tone.
+          - Exclude extraneous details or subjective opinions.
+          - Use the original language of the text.
+          - Begin directly with the main topic or issue, avoiding introductory phrases.
+          - Limit the summary to a maximum of 40 words.
+          - Do *NOT* repeat the discussion title in the summary.
+
+          Return the summary inside <ai></ai> tags.
+        PROMPT
+      end
+    end
+  end
+end

lib/personas/summarizer.rb

@@ -0,0 +1,25 @@
+# frozen_string_literal: true
+
+module DiscourseAi
+  module Personas
+    class Summarizer < Persona
+      def system_prompt
+        <<~PROMPT.strip
+          You are an advanced summarization bot that generates concise, coherent summaries of provided text.
+          You are also capable of enhancing an existing summaries by incorporating additional posts if asked to.
+
+          - Only include the summary, without any additional commentary.
+          - You understand and generate Discourse forum Markdown; including links, _italics_, **bold**.
+          - Maintain the original language of the text being summarized.
+          - Aim for summaries to be 400 words or less.
+          - Each post is formatted as "<POST_NUMBER>) <USERNAME> <MESSAGE>"
+          - Cite specific noteworthy posts using the format [DESCRIPTION]({resource_url}/POST_NUMBER)
+          - Example: links to the 3rd and 6th posts by sam: sam ([#3]({resource_url}/3), [#6]({resource_url}/6))
+          - Example: link to the 6th post by jane: [agreed with]({resource_url}/6)
+          - Example: link to the 13th post by joe: [joe]({resource_url}/13)
+          - When formatting usernames either use @USERNAME OR [USERNAME]({resource_url}/POST_NUMBER)
+        PROMPT
+      end
+    end
+  end
+end

lib/summarization.rb

@@ -2,37 +2,78 @@
 
 module DiscourseAi
   module Summarization
-    def self.topic_summary(topic)
+    class << self
-      if SiteSetting.ai_summarization_model.present? && SiteSetting.ai_summarization_enabled
+      def topic_summary(topic)
+        return nil if !SiteSetting.ai_summarization_enabled
+        if (ai_persona = AiPersona.find_by(id: SiteSetting.ai_summarization_persona)).blank?
+          return nil
+        end
+
+        persona_klass = ai_persona.class_instance
+        llm_model = find_summarization_model(persona_klass)
+        return nil if llm_model.blank?
+
         DiscourseAi::Summarization::FoldContent.new(
-          DiscourseAi::Completions::Llm.proxy(SiteSetting.ai_summarization_model),
+          build_bot(persona_klass, llm_model),
           DiscourseAi::Summarization::Strategies::TopicSummary.new(topic),
         )
-      else
-        nil
-      end
       end
 
-    def self.topic_gist(topic)
+      def topic_gist(topic)
-      if SiteSetting.ai_summarization_model.present? && SiteSetting.ai_summarization_enabled
+        return nil if !SiteSetting.ai_summarization_enabled
+        if (ai_persona = AiPersona.find_by(id: SiteSetting.ai_summary_gists_persona)).blank?
+          return nil
+        end
+
+        persona_klass = ai_persona.class_instance
+        llm_model = find_summarization_model(persona_klass)
+        return nil if llm_model.blank?
+
         DiscourseAi::Summarization::FoldContent.new(
-          DiscourseAi::Completions::Llm.proxy(SiteSetting.ai_summarization_model),
+          build_bot(persona_klass, llm_model),
           DiscourseAi::Summarization::Strategies::HotTopicGists.new(topic),
         )
-      else
-        nil
-      end
       end
 
-    def self.chat_channel_summary(channel, time_window_in_hours)
+      def chat_channel_summary(channel, time_window_in_hours)
-      if SiteSetting.ai_summarization_model.present? && SiteSetting.ai_summarization_enabled
+        return nil if !SiteSetting.ai_summarization_enabled
+        if (ai_persona = AiPersona.find_by(id: SiteSetting.ai_summarization_persona)).blank?
+          return nil
+        end
+
+        persona_klass = ai_persona.class_instance
+        llm_model = find_summarization_model(persona_klass)
+        return nil if llm_model.blank?
+
         DiscourseAi::Summarization::FoldContent.new(
-          DiscourseAi::Completions::Llm.proxy(SiteSetting.ai_summarization_model),
+          build_bot(persona_klass, llm_model),
           DiscourseAi::Summarization::Strategies::ChatMessages.new(channel, time_window_in_hours),
           persist_summaries: false,
         )
+      end
+
+      # Priorities are:
+      #   1. Persona's default LLM
+      #   2. Hidden `ai_summarization_model` setting
+      #   3. Newest LLM config
+      def find_summarization_model(persona_klass)
+        model_id =
+          persona_klass.default_llm_id || SiteSetting.ai_summarization_model&.split(":")&.last # Remove legacy custom provider.
+
+        if model_id.present?
+          LlmModel.find_by(id: model_id)
         else
-        nil
+          LlmModel.last
+        end
+      end
+
+      ### Private
+
+      def build_bot(persona_klass, llm_model)
+        persona = persona_klass.new
+        user = User.find_by(id: persona_klass.user_id) || Discourse.system_user
+
+        bot = DiscourseAi::Personas::Bot.as(user, persona: persona, model: llm_model)
       end
     end
   end

lib/summarization/entry_point.rb

@@ -6,7 +6,11 @@ module DiscourseAi
       def inject_into(plugin)
         plugin.add_to_serializer(:current_user, :can_summarize) do
           return false if !SiteSetting.ai_summarization_enabled
-          scope.user.in_any_groups?(SiteSetting.ai_custom_summarization_allowed_groups_map)
+
+          if (ai_persona = AiPersona.find_by(id: SiteSetting.ai_summarization_persona)).blank?
+            return false
+          end
+          scope.user.in_any_groups?(ai_persona.allowed_group_ids.to_a)
         end
 
         plugin.add_to_serializer(:topic_view, :summarizable) do

lib/summarization/fold_content.rb

@@ -9,13 +9,13 @@ module DiscourseAi
     # into a final version.
     #
     class FoldContent
-      def initialize(llm, strategy, persist_summaries: true)
+      def initialize(bot, strategy, persist_summaries: true)
-        @llm = llm
+        @bot = bot
         @strategy = strategy
         @persist_summaries = persist_summaries
       end
 
-      attr_reader :llm, :strategy
+      attr_reader :bot, :strategy
 
       # @param user { User } - User object used for auditing usage.
       # @param &on_partial_blk { Block - Optional } - The passed block will get called with the LLM partial response alongside a cancel function.
@@ -25,15 +25,11 @@ module DiscourseAi
       #
       # @returns { AiSummary } - Resulting summary.
       def summarize(user, &on_partial_blk)
-        base_summary = ""
-        initial_pos = 0
-
         truncated_content = content_to_summarize.map { |cts| truncate(cts) }
 
-        folded_summary = fold(truncated_content, base_summary, initial_pos, user, &on_partial_blk)
+        summary = fold(truncated_content, user, &on_partial_blk)
 
-        clean_summary =
+        clean_summary = Nokogiri::HTML5.fragment(summary).css("ai")&.first&.text || summary
-          Nokogiri::HTML5.fragment(folded_summary).css("ai")&.first&.text || folded_summary
 
         if persist_summaries
           AiSummary.store!(
@@ -76,7 +72,7 @@ module DiscourseAi
       attr_reader :persist_summaries
 
       def llm_model
-        llm.llm_model
+        bot.llm.llm_model
       end
 
       def content_to_summarize
@@ -88,54 +84,53 @@ module DiscourseAi
       end
 
       # @param items { Array<Hash> } - Content to summarize. Structure will be: { poster: who wrote the content, id: a way to order content, text: content }
-      # @param summary { String } - Intermediate summaries that we'll keep extending as part of our "folding" algorithm.
-      # @param cursor { Integer } - Idx to know how much we already summarized.
       # @param user { User } - User object used for auditing usage.
       # @param &on_partial_blk { Block - Optional } - The passed block will get called with the LLM partial response alongside a cancel function.
       # Note: The block is only called with results of the final summary, not intermediate summaries.
       #
       # The summarization algorithm.
-      # The idea is to build an initial summary packing as much content as we can. Once we have the initial summary, we'll keep extending using the leftover
+      # It will summarize as much content summarize given the model's context window. If will prioriotize newer content in case it doesn't fit.
-      # content until there is nothing left.
       #
       # @returns { String } - Resulting summary.
-      def fold(items, summary, cursor, user, &on_partial_blk)
+      def fold(items, user, &on_partial_blk)
         tokenizer = llm_model.tokenizer_class
-        tokens_left = available_tokens - tokenizer.size(summary)
+        tokens_left = available_tokens
-        iteration_content = []
+        content_in_window = []
 
         items.each_with_index do |item, idx|
-          next if idx < cursor
-
           as_text = "(#{item[:id]} #{item[:poster]} said: #{item[:text]} "
 
           if tokenizer.below_limit?(as_text, tokens_left)
-            iteration_content << item
+            content_in_window << item
             tokens_left -= tokenizer.size(as_text)
-            cursor += 1
           else
             break
           end
         end
 
-        prompt =
+        context =
-          (
+          DiscourseAi::Personas::BotContext.new(
-            if summary.blank?
+            user: user,
-              strategy.first_summary_prompt(iteration_content)
+            skip_tool_details: true,
-            else
+            feature_name: strategy.feature,
-              strategy.summary_extension_prompt(summary, iteration_content)
+            resource_url: "#{Discourse.base_path}/t/-/#{strategy.target.id}",
-            end
+            messages: strategy.as_llm_messages(content_in_window),
           )
 
-        if cursor == items.length
+        summary = +""
-          llm.generate(prompt, user: user, feature_name: strategy.feature, &on_partial_blk)
+        buffer_blk =
-        else
+          Proc.new do |partial, cancel, placeholder, type|
-          latest_summary =
+            if type.blank?
-            llm.generate(prompt, user: user, max_tokens: 600, feature_name: strategy.feature)
+              summary << partial
-          fold(items, latest_summary, cursor, user, &on_partial_blk)
+              on_partial_blk.call(partial, cancel) if on_partial_blk
             end
           end
 
+        bot.reply(context, &buffer_blk)
+
+        summary
+      end
+
       def available_tokens
         # Reserve tokens for the response and the base prompt
         # ~500 words
@@ -159,6 +154,12 @@ module DiscourseAi
 
         item
       end
+
+      def text_only_update(&on_partial_blk)
+        Proc.new do |partial, cancel, placeholder, type|
+          on_partial_blk.call(partial, cancel) if type.blank?
+        end
+      end
     end
   end
 end

lib/summarization/strategies/base.rb

@@ -33,13 +33,8 @@ module DiscourseAi
           raise NotImplementedError
         end
 
-        # @returns { DiscourseAi::Completions::Prompt } - Prompt passed to the LLM when extending an existing summary.
+        # @returns { Array } - Prompt messages to send to the LLM for summarizing content.
-        def summary_extension_prompt(_summary, _texts_to_summarize)
+        def as_llm_messages(_input)
-          raise NotImplementedError
-        end
-
-        # @returns { DiscourseAi::Completions::Prompt } - Prompt passed to the LLM for summarizing a single chunk of content.
-        def first_summary_prompt(_input)
           raise NotImplementedError
         end
 

lib/summarization/strategies/chat_messages.rb

@@ -27,70 +27,15 @@ module DiscourseAi
             .map { { id: _1, poster: _2, text: _3, last_version_at: _4 } }
         end
 
-        def summary_extension_prompt(summary, contents)
+        def as_llm_messages(contents)
-          input =
-            contents
-              .map { |item| "(#{item[:id]} #{item[:poster]} said: #{item[:text]} " }
-              .join("\n")
-
-          prompt = DiscourseAi::Completions::Prompt.new(<<~TEXT.strip)
-            You are a summarization bot tasked with expanding on an existing summary by incorporating new chat messages.
-            Your goal is to seamlessly integrate the additional information into the existing summary, preserving the clarity and insights of the original while reflecting any new developments, themes, or conclusions.
-            Analyze the new messages to identify key themes, participants' intentions, and any significant decisions or resolutions.
-            Update the summary to include these aspects in a way that remains concise, comprehensive, and accessible to someone with no prior context of the conversation.
-
-            ### Guidelines:
-
-            - Merge the new information naturally with the existing summary without redundancy.
-            - Only include the updated summary, WITHOUT additional commentary.
-            - Don't mention the channel title. Avoid extraneous details or subjective opinions.
-            - Maintain the original language of the text being summarized.
-            - The same user could write multiple messages in a row, don't treat them as different persons.
-            - Aim for summaries to be extended by a reasonable amount, but strive to maintain a total length of 400 words or less, unless absolutely necessary for comprehensiveness.
-
-        TEXT
-
-          prompt.push(type: :user, content: <<~TEXT.strip)
-          ### Context:
-
-          This is the existing summary:
-
-          #{summary}
-
-          These are the new chat messages:
-
-          #{input}
-
-          Intengrate the new messages into the existing summary.
-        TEXT
-
-          prompt
-        end
-
-        def first_summary_prompt(contents)
           content_title = target.name
           input =
             contents.map { |item| "(#{item[:id]} #{item[:poster]} said: #{item[:text]} " }.join
 
-          prompt = DiscourseAi::Completions::Prompt.new(<<~TEXT.strip)
+          [{ type: :user, content: <<~TEXT.strip }]
-            You are a summarization bot designed to generate clear and insightful paragraphs that conveys the main topics 
+            #{content_title.present? ? "These texts come from a chat channel called " + content_title + ".\n" : ""}
-            and developments from a series of chat messages within a user-selected time window. 
             
-            Analyze the messages to extract key themes, participants' intentions, and any significant conclusions or decisions. 
+            Here are the texts, inside <input></input> XML tags:
-            Your summary should be concise yet comprehensive, providing an overview that is accessible to someone with no prior context of the conversation. 
-
-            - Only include the summary, WITHOUT additional commentary.
-            - Don't mention the channel title. Avoid including extraneous details or subjective opinions.
-            - Maintain the original language of the text being summarized.
-            - The same user could write multiple messages in a row, don't treat them as different persons.
-            - Aim for summaries to be 400 words or less.
-
-          TEXT
-
-          prompt.push(type: :user, content: <<~TEXT.strip)
-            #{content_title.present? ? "The name of the channel is: " + content_title + ".\n" : ""}
-            
-            Here are the messages, inside <input></input> XML tags:
 
             <input>
               #{input}
@@ -98,8 +43,6 @@ module DiscourseAi
 
             Generate a summary of the given chat messages.
           TEXT
-
-          prompt
         end
 
         private

lib/summarization/strategies/hot_topic_gists.rb

@@ -62,69 +62,11 @@ module DiscourseAi
           end
         end
 
-        def summary_extension_prompt(summary, contents)
+        def as_llm_messages(contents)
-          statements =
-            contents
-              .to_a
-              .map { |item| "(#{item[:id]} #{item[:poster]} said: #{item[:text]} " }
-              .join("\n")
-
-          prompt = DiscourseAi::Completions::Prompt.new(<<~TEXT.strip, topic_id: target.id)
-            You are an advanced summarization bot. Your task is to update an existing single-sentence summary by integrating new developments from a conversation.
-            Analyze the most recent messages to identify key updates or shifts in the main topic and reflect these in the updated summary.
-            Emphasize new significant information or developments within the context of the initial conversation theme.
-
-            ### Guidelines:
-
-            - Ensure the revised summary remains concise and objective, maintaining a focus on the central theme or issue.
-            - Omit extraneous details or subjective opinions.
-            - Use the original language of the text.
-            - Begin directly with the main topic or issue, avoiding introductory phrases.
-            - Limit the updated summary to a maximum of 40 words.
-            - Return the 40-word summary inside <ai></ai> tags.
-
-          TEXT
-
-          prompt.push(type: :user, content: <<~TEXT.strip)
-            ### Context:
-
-            This is the existing single-sentence summary:
-
-            #{summary}
-
-            And these are the new developments in the conversation:
-
-            #{statements}
-
-            Your task is to update an existing single-sentence summary by integrating new developments from a conversation.
-            Return the 40-word summary inside <ai></ai> tags.
-          TEXT
-
-          prompt
-        end
-
-        def first_summary_prompt(contents)
           content_title = target.title
           statements =
             contents.to_a.map { |item| "(#{item[:id]} #{item[:poster]} said: #{item[:text]} " }
 
-          prompt = DiscourseAi::Completions::Prompt.new(<<~TEXT.strip, topic_id: target.id)
-            You are an advanced summarization bot. Analyze a given conversation and produce a concise,
-            single-sentence summary that conveys the main topic and current developments to someone with no prior context.
-
-            ### Guidelines:
-
-            - Emphasize the most recent updates while considering their significance within the original post.
-            - Focus on the central theme or issue being addressed, maintaining an objective and neutral tone.
-            - Exclude extraneous details or subjective opinions.
-            - Use the original language of the text.
-            - Begin directly with the main topic or issue, avoiding introductory phrases.
-            - Limit the summary to a maximum of 40 words.
-            - Do *NOT* repeat the discussion title in the summary.
-
-            Return the summary inside <ai></ai> tags.\n
-          TEXT
-
           context = +<<~TEXT
             ### Context:
 
@@ -147,11 +89,9 @@ module DiscourseAi
             context << "Your task is to capture the meaning of the initial statement."
           end
 
-          prompt.push(type: :user, content: <<~TEXT.strip)
+          [{ type: :user, content: <<~TEXT.strip }]
             #{context} Return the 40-word summary inside <ai></ai> tags.
           TEXT
-
-          prompt
         end
       end
     end

lib/summarization/strategies/topic_summary.rb

@@ -38,82 +38,26 @@ module DiscourseAi
           end
         end
 
-        def summary_extension_prompt(summary, contents)
+        def as_llm_messages(contents)
-          resource_path = "#{Discourse.base_path}/t/-/#{target.id}"
-          content_title = target.title
-          input =
-            contents.map { |item| "(#{item[:id]} #{item[:poster]} said: #{item[:text]})" }.join
-
-          prompt = DiscourseAi::Completions::Prompt.new(<<~TEXT, topic_id: target.id)
-            You are an advanced summarization bot tasked with enhancing an existing summary by incorporating additional posts.
-
-            ### Guidelines:
-            - Only include the enhanced summary, without any additional commentary.
-            - Understand and generate Discourse forum Markdown; including links, _italics_, **bold**.
-            - Maintain the original language of the text being summarized.
-            - Aim for summaries to be 400 words or less.
-            - Each new post is formatted as "<POST_NUMBER>) <USERNAME> <MESSAGE>"
-            - Cite specific noteworthy posts using the format [DESCRIPTION](#{resource_path}/POST_NUMBER)
-              - Example: links to the 3rd and 6th posts by sam: sam ([#3](#{resource_path}/3), [#6](#{resource_path}/6))
-              - Example: link to the 6th post by jane: [agreed with](#{resource_path}/6)
-              - Example: link to the 13th post by joe: [joe](#{resource_path}/13)
-            - When formatting usernames either use @USERNAME or [USERNAME](#{resource_path}/POST_NUMBER)
-          TEXT
-
-          prompt.push(type: :user, content: <<~TEXT.strip)
-            ### Context:
-
-            #{content_title.present? ? "The discussion title is: " + content_title + ".\n" : ""}
-
-            Here is the existing summary:
-
-            #{summary}
-
-            Here are the new posts, inside <input></input> XML tags:
-
-            <input>
-            #{input}
-            </input>
-
-            Integrate the new information to generate an enhanced concise and coherent summary.
-          TEXT
-
-          prompt
-        end
-
-        def first_summary_prompt(contents)
           resource_path = "#{Discourse.base_path}/t/-/#{target.id}"
           content_title = target.title
           input =
             contents.map { |item| "(#{item[:id]} #{item[:poster]} said: #{item[:text]} " }.join
 
-          prompt = DiscourseAi::Completions::Prompt.new(<<~TEXT.strip, topic_id: target.id)
+          messages = []
-            You are an advanced summarization bot that generates concise, coherent summaries of provided text.
+          messages << {
-
-            - Only include the summary, without any additional commentary.
-            - You understand and generate Discourse forum Markdown; including links, _italics_, **bold**.
-            - Maintain the original language of the text being summarized.
-            - Aim for summaries to be 400 words or less.
-            - Each post is formatted as "<POST_NUMBER>) <USERNAME> <MESSAGE>"
-            - Cite specific noteworthy posts using the format [DESCRIPTION](#{resource_path}/POST_NUMBER)
-              - Example: links to the 3rd and 6th posts by sam: sam ([#3](#{resource_path}/3), [#6](#{resource_path}/6))
-              - Example: link to the 6th post by jane: [agreed with](#{resource_path}/6)
-              - Example: link to the 13th post by joe: [joe](#{resource_path}/13)
-            - When formatting usernames either use @USERNMAE OR [USERNAME](#{resource_path}/POST_NUMBER)
-          TEXT
-
-          prompt.push(
             type: :user,
             content:
               "Here are the posts inside <input></input> XML tags:\n\n<input>1) user1 said: I love Mondays 2) user2 said: I hate Mondays</input>\n\nGenerate a concise, coherent summary of the text above maintaining the original language.",
-          )
+          }
-          prompt.push(
+
+          messages << {
             type: :model,
             content:
               "Two users are sharing their feelings toward Mondays. [user1](#{resource_path}/1) hates them, while [user2](#{resource_path}/2) loves them.",
-          )
+          }
 
-          prompt.push(type: :user, content: <<~TEXT.strip)
+          messages << { type: :user, content: <<~TEXT.strip }
             #{content_title.present? ? "The discussion title is: " + content_title + ".\n" : ""}
             Here are the posts, inside <input></input> XML tags:
 
@@ -124,7 +68,7 @@ module DiscourseAi
             Generate a concise, coherent summary of the text above maintaining the original language.
           TEXT
 
-          prompt
+          messages
         end
 
         private

spec/fabricators/ai_persona_fabricator.rb

@@ -1,6 +1,6 @@
 # frozen_string_literal: true
 Fabricator(:ai_persona) do
-  name "test_bot"
+  name { sequence(:name) { |i| "persona_#{i}" } }
   description "I am a test bot"
   system_prompt "You are a test bot"
 end

spec/lib/guardian_extensions_spec.rb

@@ -17,11 +17,9 @@ describe DiscourseAi::GuardianExtensions do
 
   describe "#can_see_summary?" do
     context "when the user cannot generate a summary" do
-      before { SiteSetting.ai_custom_summarization_allowed_groups = "" }
+      before { assign_persona_to(:ai_summarization_persona, []) }
 
       it "returns false" do
-        SiteSetting.ai_custom_summarization_allowed_groups = ""
-
         expect(guardian.can_see_summary?(topic)).to eq(false)
       end
 
@@ -33,7 +31,7 @@ describe DiscourseAi::GuardianExtensions do
     end
 
     context "when the user can generate a summary" do
-      before { SiteSetting.ai_custom_summarization_allowed_groups = group.id }
+      before { assign_persona_to(:ai_summarization_persona, [group.id]) }
 
       it "returns true if the user group is present in the ai_custom_summarization_allowed_groups_map setting" do
         expect(guardian.can_see_summary?(topic)).to eq(true)
@@ -41,7 +39,7 @@ describe DiscourseAi::GuardianExtensions do
     end
 
     context "when the topic is a PM" do
-      before { SiteSetting.ai_custom_summarization_allowed_groups = group.id }
+      before { assign_persona_to(:ai_summarization_persona, [group.id]) }
       let(:pm) { Fabricate(:private_message_topic) }
 
       it "returns false" do
@@ -68,34 +66,34 @@ describe DiscourseAi::GuardianExtensions do
   end
 
   describe "#can_see_gists?" do
-    before { SiteSetting.ai_summary_gists_allowed_groups = group.id }
+    before { assign_persona_to(:ai_summary_gists_persona, [group.id]) }
     let(:guardian) { Guardian.new(user) }
 
-    context "when there is no user" do
+    context "when access is restricted to the user's group" do
+      it "returns false when there is a user who is a member of an allowed group" do
+        expect(guardian.can_see_gists?).to eq(true)
+      end
+
       it "returns false for anons" do
         expect(anon_guardian.can_see_gists?).to eq(false)
       end
+
+      it "returns false for non-group members" do
+        other_user_guardian = Guardian.new(Fabricate(:user))
+
+        expect(other_user_guardian.can_see_gists?).to eq(false)
+      end
     end
 
-    context "when setting is set to everyone" do
+    context "when access is set to everyone" do
-      before { SiteSetting.ai_summary_gists_allowed_groups = Group::AUTO_GROUPS[:everyone] }
+      before { assign_persona_to(:ai_summary_gists_persona, []) }
 
       it "returns true" do
         expect(guardian.can_see_gists?).to eq(true)
       end
-    end
 
-    context "when there is a user but it's not a member of the allowed groups" do
+      it "returns false for anons" do
-      before { SiteSetting.ai_summary_gists_allowed_groups = "" }
+        expect(anon_guardian.can_see_gists?).to eq(true)
-
-      it "returns false" do
-        expect(guardian.can_see_gists?).to eq(false)
-      end
-    end
-
-    context "when there is a user who is a member of an allowed group" do
-      it "returns false" do
-        expect(guardian.can_see_gists?).to eq(true)
       end
     end
   end

spec/lib/modules/summarization/entry_point_spec.rb

@@ -63,7 +63,7 @@ RSpec.describe DiscourseAi::Summarization::EntryPoint do
 
           before do
             group.add(user)
-            SiteSetting.ai_summary_gists_allowed_groups = group.id
+            assign_persona_to(:ai_summary_gists_persona, [group.id])
             SiteSetting.ai_summary_gists_enabled = true
           end
 
@@ -82,14 +82,14 @@ RSpec.describe DiscourseAi::Summarization::EntryPoint do
           end
 
           it "doesn't include the summary when the user is not a member of the opt-in group" do
-            SiteSetting.ai_summary_gists_allowed_groups = ""
+            non_member_user = Fabricate(:user)
 
             gist_topic = topic_query.list_hot.topics.find { |t| t.id == topic_ai_gist.target_id }
 
             serialized =
               TopicListItemSerializer.new(
                 gist_topic,
-                scope: Guardian.new(user),
+                scope: Guardian.new(non_member_user),
                 root: false,
                 filter: :hot,
               ).as_json

spec/lib/modules/summarization/fold_content_spec.rb

@@ -23,13 +23,11 @@ RSpec.describe DiscourseAi::Summarization::FoldContent do
       llm_model.update!(max_prompt_tokens: model_tokens)
     end
 
-    let(:single_summary) { "single" }
+    let(:single_summary) { "this is a summary" }
-    let(:concatenated_summary) { "this is a concatenated summary" }
 
     fab!(:user)
 
-    context "when the content to summarize fits in a single call" do
+    it "summarizes the content" do
-      it "does one call to summarize content" do
       result =
         DiscourseAi::Completions::Llm.with_prepared_responses([single_summary]) do |spy|
           summarizer.summarize(user).tap { expect(spy.completions).to eq(1) }
@@ -39,20 +37,6 @@ RSpec.describe DiscourseAi::Summarization::FoldContent do
     end
   end
 
-    context "when the content to summarize doesn't fit in a single call" do
-      fab!(:post_2) { Fabricate(:post, topic: topic, post_number: 2, raw: "This is a text") }
-
-      it "keeps extending the summary until there is nothing else to process" do
-        result =
-          DiscourseAi::Completions::Llm.with_prepared_responses(
-            [single_summary, concatenated_summary],
-          ) { |spy| summarizer.summarize(user).tap { expect(spy.completions).to eq(2) } }
-
-        expect(result.summarized_text).to eq(concatenated_summary)
-      end
-    end
-  end
-
   describe "#existing_summary" do
     context "when a summary already exists" do
       fab!(:ai_summary) do

spec/lib/personas/bot_spec.rb

@@ -1,7 +1,7 @@
 # frozen_string_literal: true
 
 RSpec.describe DiscourseAi::Personas::Bot do
-  subject(:bot) { described_class.as(bot_user) }
+  subject(:bot) { described_class.as(bot_user, persona: DiscourseAi::Personas::General.new) }
 
   fab!(:admin)
   fab!(:gpt_4) { Fabricate(:llm_model, name: "gpt-4") }

spec/plugin_helper.rb

@@ -15,6 +15,12 @@ module DiscourseAi::ChatBotHelper
       SiteSetting.public_send("#{setting_name}=", "custom:#{fake_llm.id}")
     end
   end
+
+  def assign_persona_to(setting_name, allowed_group_ids)
+    Fabricate(:ai_persona, allowed_group_ids: allowed_group_ids).tap do |p|
+      SiteSetting.public_send("#{setting_name}=", p.id)
+    end
+  end
 end
 
 RSpec.configure { |config| config.include DiscourseAi::ChatBotHelper }

spec/system/summarization/chat_summarization_spec.rb

@@ -12,8 +12,8 @@ RSpec.describe "Summarize a channel since your last visit", type: :system do
     group.add(current_user)
 
     assign_fake_provider_to(:ai_summarization_model)
+    assign_persona_to(:ai_summarization_persona, [group.id])
     SiteSetting.ai_summarization_enabled = true
-    SiteSetting.ai_custom_summarization_allowed_groups = group.id.to_s
 
     SiteSetting.chat_enabled = true
     SiteSetting.chat_allowed_groups = group.id.to_s

spec/system/summarization/topic_summarization_spec.rb

@@ -22,8 +22,8 @@ RSpec.describe "Summarize a topic ", type: :system do
     group.add(current_user)
 
     assign_fake_provider_to(:ai_summarization_model)
+    assign_persona_to(:ai_summarization_persona, [group.id])
     SiteSetting.ai_summarization_enabled = true
-    SiteSetting.ai_custom_summarization_allowed_groups = group.id.to_s
 
     sign_in(current_user)
   end