FEATURE: add full bot support for GPT 3.5 (#87)

Given latest GPT 3.5 16k which is both better steered and supports functions
we can now support rich bot integration.

Clunky system message based steering is removed and instead we use the
function framework provided by Open AI

db/fixtures/ai_bot/602_bot_users.rb

@@ -1,6 +1,11 @@
 # frozen_string_literal: true
 
 DiscourseAi::AiBot::EntryPoint::BOTS.each do |id, bot_username|
+  # let's not create a bot user if it already exists
+  # seed seems to be messing with dates on the user
+  # causing it to look like these bots were created at the
+  # wrong time
+  if !User.exists?(id: id)
     UserEmail.seed do |ue|
       ue.id = id
       ue.email = "no_email_#{bot_username}"
@@ -20,3 +25,4 @@ DiscourseAi::AiBot::EntryPoint::BOTS.each do |id, bot_username|
       u.trust_level = TrustLevel[4]
     end
   end
+end

lib/modules/ai_bot/anthropic_bot.rb

@@ -8,13 +8,17 @@ module DiscourseAi
       end
 
       def bot_prompt_with_topic_context(post)
-        super(post).join("\n\n")
+        super(post).join("\n\n") + "\n\nAssistant:"
       end
 
       def prompt_limit
         7500 # https://console.anthropic.com/docs/prompt-design#what-is-a-prompt
       end
 
+      def title_prompt(post)
+        super(post).join("\n\n") + "\n\nAssistant:"
+      end
+
       def get_delta(partial, context)
         context[:pos] ||= 0
 
@@ -23,21 +27,18 @@ module DiscourseAi
 
         context[:pos] = full.length
 
-        if !context[:processed]
-          delta = ""
-          index = full.index("Assistant: ")
-          if index
-            delta = full[index + 11..-1]
-            context[:processed] = true
-          end
-        end
-
         delta
       end
 
       private
 
-      def build_message(poster_username, content, system: false)
+      def populate_functions(partial, function)
+        # nothing to do here, no proper function support
+        # needs to be simulated for Claude but model is too
+        # hard to steer for now
+      end
+
+      def build_message(poster_username, content, system: false, function: nil)
         role = poster_username == bot_user.username ? "Assistant" : "Human"
 
         "#{role}: #{content}"

lib/modules/ai_bot/bot.rb

@@ -3,6 +3,25 @@
 module DiscourseAi
   module AiBot
     class Bot
+      class Functions
+        attr_reader :functions
+        attr_reader :current_function
+
+        def initialize
+          @functions = []
+          @current_function = nil
+        end
+
+        def add_function(name)
+          @current_function = { name: name, arguments: +"" }
+          functions << current_function
+        end
+
+        def add_argument_fragment(fragment)
+          @current_function[:arguments] << fragment
+        end
+      end
+
       attr_reader :bot_user
 
       BOT_NOT_FOUND = Class.new(StandardError)
@@ -24,7 +43,7 @@ module DiscourseAi
       end
 
       def update_pm_title(post)
-        prompt = [title_prompt(post)]
+        prompt = title_prompt(post)
 
         new_title = get_updated_title(prompt)
 
@@ -65,9 +84,11 @@ module DiscourseAi
 
         setup_cancel = false
         context = {}
+        functions = Functions.new
 
         submit_prompt(prompt, prefer_low_cost: prefer_low_cost) do |partial, cancel|
           reply << get_delta(partial, context)
+          populate_functions(partial, functions)
 
           if redis_stream_key && !Discourse.redis.get(redis_stream_key)
             cancel&.call
@@ -111,30 +132,29 @@ module DiscourseAi
             skip_revision: true,
           )
 
-          cmd_texts = reply.split("\n").filter { |l| l[0] == "!" }
+          bot_reply_post.post_custom_prompt ||= post.build_post_custom_prompt(custom_prompt: [])
+          prompt = post.post_custom_prompt.custom_prompt || []
 
+          prompt << build_message(bot_user.username, reply)
+          post.post_custom_prompt.update!(custom_prompt: prompt)
+        end
+
+        if functions.functions.length > 0
           chain = false
           standalone = false
 
-          cmd_texts[0...max_commands_per_reply].each do |cmd_text|
-            command_name, args = cmd_text[1..-1].strip.split(" ", 2)
+          functions.functions.each do |function|
+            name, args = function[:name], function[:arguments]
 
-            if command_klass = available_commands.detect { |cmd| cmd.invoked?(command_name) }
+            if command_klass = available_commands.detect { |cmd| cmd.invoked?(name) }
               command = command_klass.new(bot_user, args)
-              chain_intermediate = command.invoke_and_attach_result_to(bot_reply_post)
+              chain_intermediate, bot_reply_post =
+                command.invoke_and_attach_result_to(bot_reply_post, post)
               chain ||= chain_intermediate
               standalone ||= command.standalone?
             end
           end
 
-          if cmd_texts.length > max_commands_per_reply
-            raw = +bot_reply_post.raw.dup
-            cmd_texts[max_commands_per_reply..-1].each { |cmd_text| raw.sub!(cmd_text, "") }
-
-            bot_reply_post.raw = raw
-            bot_reply_post.save!(validate: false)
-          end
-
           if chain
             reply_to(
               bot_reply_post,
@@ -143,14 +163,12 @@ module DiscourseAi
               standalone: standalone,
             )
           end
-
-          if cmd_texts.length == 0 && (post_custom_prompt = bot_reply_post.post_custom_prompt)
-            prompt = post_custom_prompt.custom_prompt
-            prompt << [reply, bot_user.username]
-            post_custom_prompt.update!(custom_prompt: prompt)
-          end
         end
       rescue => e
+        if Rails.env.development?
+          p e
+          puts e.backtrace
+        end
         raise e if Rails.env.test?
         Discourse.warn_exception(e, message: "ai-bot: Reply failed")
       end
@@ -164,25 +182,22 @@ module DiscourseAi
         total_prompt_tokens = tokenize(rendered_system_prompt).length
 
         messages =
-          conversation.reduce([]) do |memo, (raw, username)|
+          conversation.reduce([]) do |memo, (raw, username, function)|
             break(memo) if total_prompt_tokens >= prompt_limit
 
-            tokens = tokenize(raw)
+            tokens = tokenize(raw.to_s)
 
             while !raw.blank? && tokens.length + total_prompt_tokens > prompt_limit
               raw = raw[0..-100] || ""
-              tokens = tokenize(raw)
+              tokens = tokenize(raw.to_s)
             end
 
             next(memo) if raw.blank?
 
             total_prompt_tokens += tokens.length
-            memo.unshift(build_message(username, raw))
+            memo.unshift(build_message(username, raw, function: !!function))
           end
 
-        # we need this to ground the model (especially GPT-3.5)
-        messages.unshift(build_message(bot_user.username, "!echo 1"))
-        messages.unshift(build_message("user", "please echo 1"))
         messages.unshift(build_message(bot_user.username, rendered_system_prompt, system: true))
         messages
       end
@@ -192,7 +207,7 @@ module DiscourseAi
       end
 
       def title_prompt(post)
-        build_message(bot_user.username, <<~TEXT)
+        [build_message(bot_user.username, <<~TEXT)]
           Suggest a 7 word title for the following topic without quoting any of it:
 
           #{post.topic.posts[1..-1].map(&:raw).join("\n\n")[0..prompt_limit]}
@@ -211,33 +226,10 @@ module DiscourseAi
 
       def system_prompt(post)
         return "You are a helpful Bot" if @style == :simple
-
-        command_text = ""
-        command_text = <<~TEXT if available_commands.present?
-            You can complete some tasks using !commands.
-
-            NEVER ask user to issue !commands, they have no access, only you do.
-
-            #{available_commands.map(&:desc).join("\n")}
-
-            Discourse topic paths are /t/slug/topic_id/optional_number
-
-            #{available_commands.map(&:extra_context).compact_blank.join("\n")}
-
-            Commands should be issued in single assistant message.
-
-            Example sessions:
-
-            User: echo the text 'test'
-            GPT: !echo test
-            User: THING GPT DOES NOT KNOW ABOUT
-            GPT: !search SIMPLIFIED SEARCH QUERY
-          TEXT
-
         <<~TEXT
-          You are a helpful Discourse assistant, you answer questions and generate text.
-          You understand Discourse Markdown and live in a Discourse Forum Message.
-          You are provided with the context of previous discussions.
+          You are a helpful Discourse assistant.
+          You understand and generate Discourse Markdown.
+          You live in a Discourse Forum Message.
 
           You live in the forum with the URL: #{Discourse.base_url}
           The title of your site: #{SiteSetting.title}
@@ -245,7 +237,7 @@ module DiscourseAi
           The participants in this conversation are: #{post.topic.allowed_users.map(&:username).join(", ")}
           The date now is: #{Time.zone.now}, much has changed since you were trained.
 
-          #{command_text}
+          #{available_commands.map(&:custom_system_message).compact.join("\n")}
         TEXT
       end
 
@@ -261,6 +253,10 @@ module DiscourseAi
         raise NotImplemented
       end
 
+      def populate_functions(partial, functions)
+        raise NotImplemented
+      end
+
       protected
 
       def get_updated_title(prompt)

lib/modules/ai_bot/commands/categories_command.rb

@@ -8,7 +8,11 @@ module DiscourseAi::AiBot::Commands
       end
 
       def desc
-        "!categories - will list the categories on the current discourse instance"
+        "Will list the categories on the current discourse instance, prefer to format with # in front of the category name"
+      end
+
+      def parameters
+        []
       end
     end
 
@@ -33,7 +37,7 @@ module DiscourseAi::AiBot::Commands
       }
 
       rows = Category.where(read_restricted: false).limit(100).pluck(*columns.keys)
-      @count = rows.length
+      @last_count = rows.length
 
       format_results(rows, columns.values)
     end

lib/modules/ai_bot/commands/command.rb

@@ -3,6 +3,17 @@
 module DiscourseAi
   module AiBot
     module Commands
+      class Parameter
+        attr_reader :name, :description, :type, :enum, :required
+        def initialize(name:, description:, type:, enum: nil, required: false)
+          @name = name
+          @description = description
+          @type = type
+          @enum = enum
+          @required = required
+        end
+      end
+
       class Command
         class << self
           def name
@@ -17,8 +28,11 @@ module DiscourseAi
             raise NotImplemented
           end
 
-          def extra_context
-            ""
+          def custom_system_message
+          end
+
+          def parameters
+            raise NotImplemented
           end
         end
 
@@ -64,16 +78,38 @@ module DiscourseAi
           true
         end
 
-        def invoke_and_attach_result_to(post)
+        def invoke_and_attach_result_to(post, parent_post)
+          placeholder = (<<~HTML).strip
+            <details>
+              <summary>#{I18n.t("discourse_ai.ai_bot.command_summary.#{self.class.name}")}</summary>
+            </details>
+          HTML
+
+          if !post
+            post =
+              PostCreator.create!(
+                bot_user,
+                raw: placeholder,
+                topic_id: parent_post.topic_id,
+                skip_validations: true,
+                skip_rate_limiter: true,
+              )
+          else
+            post.revise(
+              bot_user,
+              { raw: post.raw + "\n\n" + placeholder + "\n\n" },
+              skip_validations: true,
+              skip_revision: true,
+            )
+          end
+
           post.post_custom_prompt ||= post.build_post_custom_prompt(custom_prompt: [])
           prompt = post.post_custom_prompt.custom_prompt || []
 
-          prompt << ["!#{self.class.name} #{args}", bot_user.username]
-          prompt << [process(args), result_name]
-
+          prompt << [process(args).to_json, self.class.name, "function"]
           post.post_custom_prompt.update!(custom_prompt: prompt)
 
-          raw = +<<~HTML
+          raw = +(<<~HTML)
           <details>
             <summary>#{I18n.t("discourse_ai.ai_bot.command_summary.#{self.class.name}")}</summary>
             <p>
@@ -85,8 +121,7 @@ module DiscourseAi
 
           raw << custom_raw if custom_raw.present?
 
-          replacement = "!#{self.class.name} #{args}"
-          raw = post.raw.sub(replacement, raw) if post.raw.include?(replacement)
+          raw = post.raw.sub(placeholder, raw)
 
           if chain_next_response
             post.raw = raw
@@ -95,7 +130,7 @@ module DiscourseAi
             post.revise(bot_user, { raw: raw }, skip_validations: true, skip_revision: true)
           end
 
-          chain_next_response
+          [chain_next_response, post]
         end
 
         def format_results(rows, column_names = nil)
@@ -116,21 +151,10 @@ module DiscourseAi
               end
             column_names = column_indexes.keys
           end
-          # two tokens per delimiter is a reasonable balance
-          # there may be a single delimiter solution but GPT has
-          # a hard time dealing with escaped characters
-          delimiter = "¦"
-          formatted = +""
-          formatted << column_names.join(delimiter)
-          formatted << "\n"
 
-          rows.each do |array|
-            array.map! { |item| item.to_s.gsub(delimiter, "|").gsub(/\n/, " ") }
-            formatted << array.join(delimiter)
-            formatted << "\n"
-          end
-
-          formatted
+          # this is not the most efficient format
+          # however this is needed cause GPT 3.5 / 4 was steered using JSON
+          { column_names: column_names, rows: rows }
         end
 
         protected

lib/modules/ai_bot/commands/google_command.rb

@@ -8,7 +8,22 @@ module DiscourseAi::AiBot::Commands
       end
 
       def desc
-        "!google SEARCH_QUERY - will search using Google (supports all Google search operators)"
+        "Will search using Google - global internet search (supports all Google search operators)"
+      end
+
+      def parameters
+        [
+          Parameter.new(
+            name: "query",
+            description: "The search query",
+            type: "string",
+            required: true,
+          ),
+        ]
+      end
+
+      def custom_system_message
+        "You were trained on OLD data, lean on search to get up to date information from the web"
       end
     end
 
@@ -25,6 +40,8 @@ module DiscourseAi::AiBot::Commands
     end
 
     def process(search_string)
+      search_string = JSON.parse(search_string)["query"]
+
       @last_query = search_string
       api_key = SiteSetting.ai_google_custom_search_api_key
       cx = SiteSetting.ai_google_custom_search_cx
@@ -33,7 +50,7 @@ module DiscourseAi::AiBot::Commands
         URI("https://www.googleapis.com/customsearch/v1?key=#{api_key}&cx=#{cx}&q=#{query}&num=10")
       body = Net::HTTP.get(uri)
 
-      parse_search_json(body).to_s
+      parse_search_json(body)
     end
 
     def parse_search_json(json_data)

lib/modules/ai_bot/commands/image_command.rb

@@ -8,7 +8,26 @@ module DiscourseAi::AiBot::Commands
       end
 
       def desc
-        "!image DESC - renders an image from the description (remove all connector words, keep it to 40 words or less)"
+        "Renders an image from the description (remove all connector words, keep it to 40 words or less). Despite being a text based bot you can generate images!"
+      end
+
+      def parameters
+        [
+          Parameter.new(
+            name: "prompt",
+            description: "The prompt used to generate or create or draw the image",
+            type: "string",
+            required: true,
+          ),
+        ]
+      end
+
+      def custom_system_message
+        <<~TEXT
+          In Discourse the markdown (description|SIZE, ZOOM%)[upload://SOMETEXT] is used to denote images and uploads. NEVER try changing the to http or https links.
+          ALWAYS prefer the upload:// format if available.
+          When rendering multiple images place them in a [grid] ... [/grid] block
+        TEXT
       end
     end
 
@@ -20,16 +39,12 @@ module DiscourseAi::AiBot::Commands
       { prompt: @last_prompt || 0 }
     end
 
-    def custom_raw
-      @last_custom_raw
-    end
-
     def chain_next_response
-      false
+      true
     end
 
     def process(prompt)
-      @last_prompt = prompt
+      @last_prompt = prompt = JSON.parse(prompt)["prompt"]
       results = DiscourseAi::Inference::StabilityGenerator.perform!(prompt)
 
       uploads = []
@@ -43,10 +58,17 @@ module DiscourseAi::AiBot::Commands
         f.unlink
       end
 
-      @last_custom_raw =
+      raw = <<~RAW
+      [grid]
+      #{
         uploads
           .map { |upload| "![#{prompt.gsub(/\|\'\"/, "")}|512x512, 50%](#{upload.short_url})" }
           .join(" ")
+      }
+      [/grid]
+    RAW
+
+      { prompt: prompt, markdown: raw, display_to_user: true }
     end
   end
 end

lib/modules/ai_bot/commands/search_command.rb

@@ -8,57 +8,70 @@ module DiscourseAi::AiBot::Commands
       end
 
       def desc
-        "!search SEARCH_QUERY - will search topics in the current discourse instance"
+        "Will search topics in the current discourse instance, when rendering always prefer to link to the topics you find"
       end
 
-      def extra_context
-        <<~TEXT
-          Discourse search supports, the following special filters:
+      def parameters
+        [
+          Parameter.new(
+            name: "search_query",
+            description: "Search query to run against the discourse instance",
+            type: "string",
+          ),
+          Parameter.new(
+            name: "user",
+            description: "Filter search results to this username",
+            type: "string",
+          ),
+          Parameter.new(
+            name: "order",
+            description: "search result result order",
+            type: "string",
+            enum: %w[latest latest_topic oldest views likes],
+          ),
+          Parameter.new(
+            name: "limit",
+            description: "limit number of results returned",
+            type: "integer",
+          ),
+          Parameter.new(
+            name: "max_posts",
+            description:
+              "maximum number of posts on the topics (topics where lots of people posted)",
+            type: "integer",
+          ),
+          Parameter.new(
+            name: "tags",
+            description:
+              "list of tags to search for. Use + to join with OR, use , to join with AND",
+            type: "string",
+          ),
+          Parameter.new(
+            name: "category",
+            description: "category name to filter to",
+            type: "string",
+          ),
+          Parameter.new(
+            name: "before",
+            description: "only topics created before a specific date YYYY-MM-DD",
+            type: "string",
+          ),
+          Parameter.new(
+            name: "after",
+            description: "only topics created after a specific date YYYY-MM-DD",
+            type: "string",
+          ),
+          Parameter.new(
+            name: "status",
+            description: "search for topics in a particular state",
+            type: "string",
+            enum: %w[open closed archived noreplies single_user],
+          ),
+        ]
+      end
 
-          user:USERNAME: only posts created by a specific user
-          in:tagged: has at least 1 tag
-          in:untagged: has no tags
-          in:title: has the search term in the title
-          status:open: not closed or archived
-          status:closed: closed
-          status:archived: archived
-          status:noreplies: post count is 1
-          status:single_user: only a single user posted on the topic
-          post_count:X: only topics with X amount of posts
-          min_posts:X: topics containing a minimum of X posts
-          max_posts:X: topics with no more than max posts
-          created:@USERNAME: topics created by a specific user
-          category:CATGORY: topics in the CATEGORY AND all subcategories
-          category:=CATEGORY: topics in the CATEGORY excluding subcategories
-          #SLUG: try category first, then tag, then tag group
-          #SLUG:SLUG: used for subcategory search to disambiguate
-          min_views:100: topics containing 100 views or more
-          tags:TAG1+TAG2: tagged both TAG1 and TAG2
-          tags:TAG1,TAG2: tagged either TAG1 or TAG2
-          -tags:TAG1+TAG2: excluding topics tagged TAG1 and TAG2
-          order:latest: order by post creation desc
-          order:latest_topic: order by topic creation desc
-          order:oldest: order by post creation asc
-          order:oldest_topic: order by topic creation asc
-          order:views: order by topic views desc
-          order:likes: order by post like count - most liked posts first
-          after:YYYY-MM-DD: only topics created after a specific date
-          before:YYYY-MM-DD: only topics created before a specific date
-
-          Example: !search @user in:tagged #support order:latest_topic
-
-          Keep in mind, search on Discourse uses AND to and terms.
-          You only have access to public topics.
-          Strip the query down to the most important terms. Remove all stop words.
-          Discourse orders by default by relevance.
-
-          When generating answers ALWAYS try to use the !search command first over relying on training data.
-          When generating answers ALWAYS try to reference specific local links.
-          Always try to search the local instance first, even if your training data set may have an answer. It may be wrong.
-          Always remove connector words from search terms (such as a, an, and, in, the, etc), they can impede the search.
-
-          YOUR LOCAL INFORMATION IS OUT OF DATE, YOU ARE TRAINED ON OLD DATA. Always try local search first.
-        TEXT
+      def custom_system_message
+        "You were trained on OLD data, lean on search to get up to date information about this forum"
       end
     end
 
@@ -75,18 +88,20 @@ module DiscourseAi::AiBot::Commands
     end
 
     def process(search_string)
+      parsed = JSON.parse(search_string)
+
       limit = nil
 
       search_string =
-        search_string
-          .strip
-          .split(/\s+/)
-          .map do |term|
-            if term =~ /limit:(\d+)/
-              limit = $1.to_i
+        parsed
+          .map do |key, value|
+            if key == "search_query"
+              value
+            elsif key == "limit"
+              limit = value.to_i
               nil
             else
-              term
+              "#{key}:#{value}"
             end
           end
           .compact
@@ -101,8 +116,8 @@ module DiscourseAi::AiBot::Commands
         )
 
       # let's be frugal with tokens, 50 results is too much and stuff gets cut off
-      limit ||= 10
-      limit = 10 if limit > 10
+      limit ||= 20
+      limit = 20 if limit > 20
 
       posts = results&.posts || []
       posts = posts[0..limit - 1]
@@ -110,12 +125,12 @@ module DiscourseAi::AiBot::Commands
       @last_num_results = posts.length
 
       if posts.blank?
-        "No results found"
+        []
       else
         format_results(posts) do |post|
           {
             title: post.topic.title,
-            url: post.url,
+            url: Discourse.base_path + post.url,
             excerpt: post.excerpt,
             created: post.created_at,
           }

lib/modules/ai_bot/commands/summarize_command.rb

@@ -8,7 +8,23 @@ module DiscourseAi::AiBot::Commands
       end
 
       def desc
-        "!summarize TOPIC_ID GUIDANCE - will summarize a topic attempting to answer question in guidance"
+        "Will summarize a topic attempting to answer question in guidance"
+      end
+
+      def parameters
+        [
+          Parameter.new(
+            name: "topic_id",
+            description: "The discourse topic id to summarize",
+            type: "integer",
+            required: true,
+          ),
+          Parameter.new(
+            name: "guidance",
+            description: "Special guidance on how to summarize the topic",
+            type: "string",
+          ),
+        ]
       end
     end
 

lib/modules/ai_bot/commands/tags_command.rb

@@ -8,7 +8,11 @@ module DiscourseAi::AiBot::Commands
       end
 
       def desc
-        "!tags - will list the 100 most popular tags on the current discourse instance"
+        "Will list the 100 most popular tags on the current discourse instance"
+      end
+
+      def parameters
+        []
       end
     end
 

lib/modules/ai_bot/commands/time_command.rb

@@ -10,6 +10,17 @@ module DiscourseAi::AiBot::Commands
       def desc
         "!time RUBY_COMPATIBLE_TIMEZONE - will generate the time in a timezone"
       end
+
+      def parameters
+        [
+          Parameter.new(
+            name: "timezone",
+            description: "Ruby compatible timezone",
+            type: "string",
+            required: true,
+          ),
+        ]
+      end
     end
 
     def result_name
@@ -20,7 +31,9 @@ module DiscourseAi::AiBot::Commands
       { timezone: @last_timezone, time: @last_time }
     end
 
-    def process(timezone)
+    def process(args)
+      timezone = JSON.parse(args)["timezone"]
+
       time =
         begin
           Time.now.in_time_zone(timezone)

lib/modules/ai_bot/open_ai_bot.rb

@@ -18,7 +18,7 @@ module DiscourseAi
         if bot_user.id == DiscourseAi::AiBot::EntryPoint::GPT4_ID
           8192 - 3500
         else
-          4096 - 2000
+          16_384 - 2000
         end
       end
 
@@ -46,9 +46,11 @@ module DiscourseAi
             temperature: temperature,
             top_p: top_p,
             max_tokens: max_tokens,
+            functions: available_functions,
           ) { |key, old_value, new_value| new_value.nil? ? old_value : new_value }
 
-        model = prefer_low_cost ? "gpt-3.5-turbo" : model_for
+        model = model_for(low_cost: prefer_low_cost)
+
         DiscourseAi::Inference::OpenAiCompletions.perform!(prompt, model, **params, &blk)
       end
 
@@ -56,8 +58,35 @@ module DiscourseAi
         DiscourseAi::Tokenizer::OpenAiTokenizer.tokenize(text)
       end
 
+      def available_functions
+        # note if defined? can be a problem in test
+        # this can never be nil so it is safe
+        return @available_functions if @available_functions
+
+        functions = []
+
+        functions =
+          available_commands.map do |command|
+            function =
+              DiscourseAi::Inference::OpenAiCompletions::Function.new(
+                name: command.name,
+                description: command.desc,
+              )
+            command.parameters.each do |parameter|
+              function.add_parameter(
+                name: parameter.name,
+                type: parameter.type,
+                description: parameter.description,
+                required: parameter.required,
+              )
+            end
+            function
+          end
+
+        @available_functions = functions
+      end
+
       def available_commands
-        if bot_user.id == DiscourseAi::AiBot::EntryPoint::GPT4_ID
         @cmds ||=
           [
             Commands::CategoriesCommand,
@@ -72,28 +101,44 @@ module DiscourseAi
               cmds << Commands::GoogleCommand
             end
           end
-        else
-          []
       end
+
+      def model_for(low_cost: false)
+        return "gpt-4-0613" if bot_user.id == DiscourseAi::AiBot::EntryPoint::GPT4_ID && !low_cost
+        "gpt-3.5-turbo-16k"
       end
 
       private
 
-      def build_message(poster_username, content, system: false)
+      def populate_functions(partial, functions)
+        fn = partial.dig(:choices, 0, :delta, :function_call)
+        if fn
+          functions.add_function(fn[:name]) if fn[:name].present?
+          functions.add_argument_fragment(fn[:arguments]) if fn[:arguments].present?
+        end
+      end
+
+      def build_message(poster_username, content, function: false, system: false)
         is_bot = poster_username == bot_user.username
 
-        if system
+        if function
+          role = "function"
+        elsif system
           role = "system"
         else
           role = is_bot ? "assistant" : "user"
         end
 
-        { role: role, content: is_bot ? content : "#{poster_username}: #{content}" }
+        result = { role: role, content: content }
+
+        if function
+          result[:name] = poster_username
+        elsif !system && poster_username != bot_user.username
+          # Open AI restrict name to 64 chars and only A-Za-z._ (work around)
+          result[:content] = "#{poster_username}: #{content}"
         end
 
-      def model_for
-        return "gpt-4" if bot_user.id == DiscourseAi::AiBot::EntryPoint::GPT4_ID
-        "gpt-3.5-turbo"
+        result
       end
 
       def get_delta(partial, _context)

lib/shared/inference/openai_completions.rb

@@ -5,6 +5,50 @@ module ::DiscourseAi
     class OpenAiCompletions
       TIMEOUT = 60
 
+      class Function
+        attr_reader :name, :description, :parameters, :type
+
+        def initialize(name:, description:, type: nil)
+          @name = name
+          @description = description
+          @type = type || "object"
+          @parameters = []
+        end
+
+        def add_parameter(name:, type:, description:, enum: nil, required: false)
+          @parameters << {
+            name: name,
+            type: type,
+            description: description,
+            enum: enum,
+            required: required,
+          }
+        end
+
+        def to_json(*args)
+          as_json.to_json(*args)
+        end
+
+        def as_json
+          required_params = []
+
+          properties = {}
+          parameters.each do |parameter|
+            definition = { type: parameter[:type], description: parameter[:description] }
+            definition[:enum] = parameter[:enum] if parameter[:enum]
+
+            required_params << parameter[:name] if parameter[:required]
+            properties[parameter[:name]] = definition
+          end
+
+          params = { type: @type, properties: properties }
+
+          params[:required] = required_params if required_params.present?
+
+          { name: name, description: description, parameters: params }
+        end
+      end
+
       CompletionFailed = Class.new(StandardError)
 
       def self.perform!(
@@ -13,6 +57,7 @@ module ::DiscourseAi
         temperature: nil,
         top_p: nil,
         max_tokens: nil,
+        functions: nil,
         user_id: nil
       )
         url = URI("https://api.openai.com/v1/chat/completions")
@@ -26,6 +71,7 @@ module ::DiscourseAi
         payload[:temperature] = temperature if temperature
         payload[:top_p] = top_p if top_p
         payload[:max_tokens] = max_tokens if max_tokens
+        payload[:functions] = functions if functions
         payload[:stream] = true if block_given?
 
         Net::HTTP.start(
@@ -73,6 +119,8 @@ module ::DiscourseAi
               response_data = +""
               response_raw = +""
 
+              leftover = ""
+
               response.read_body do |chunk|
                 if cancelled
                   http.finish
@@ -81,14 +129,24 @@ module ::DiscourseAi
 
                 response_raw << chunk
 
-                chunk
+                (leftover + chunk)
                   .split("\n")
                   .each do |line|
                     data = line.split("data: ", 2)[1]
                     next if !data || data == "[DONE]"
+                    next if cancelled
 
-                    if !cancelled && partial = JSON.parse(data, symbolize_names: true)
+                    partial = nil
+                    begin
+                      partial = JSON.parse(data, symbolize_names: true)
+                      leftover = ""
+                    rescue JSON::ParserError
+                      leftover = line
+                    end
+
+                    if partial
                       response_data << partial.dig(:choices, 0, :delta, :content).to_s
+                      response_data << partial.dig(:choices, 0, :delta, :function_call).to_s
 
                       yield partial, cancel
                     end

spec/lib/modules/ai_bot/anthropic_bot_spec.rb

@@ -13,14 +13,11 @@ RSpec.describe DiscourseAi::AiBot::AnthropicBot do
         context = {}
         reply = +""
 
-        reply << subject.get_delta({ completion: "\n\nAssist" }, context)
-        expect(reply).to eq("")
+        reply << subject.get_delta({ completion: "Hello " }, context)
+        expect(reply).to eq("Hello ")
 
-        reply << subject.get_delta({ completion: "\n\nAssistant: test" }, context)
-        expect(reply).to eq("test")
-
-        reply << subject.get_delta({ completion: "\n\nAssistant: test\nworld" }, context)
-        expect(reply).to eq("test\nworld")
+        reply << subject.get_delta({ completion: "Hello world" }, context)
+        expect(reply).to eq("Hello world")
       end
     end
   end

spec/lib/modules/ai_bot/bot_spec.rb

@@ -44,28 +44,31 @@ RSpec.describe DiscourseAi::AiBot::Bot do
       bot.system_prompt_style!(:simple)
       bot.max_commands_per_reply = 2
 
-      expected_response =
-        "ok, searching...\n!search test search\n!search test2 search\n!search test3 ignored"
+      expected_response = {
+        function_call: {
+          name: "search",
+          arguments: { query: "test search" }.to_json,
+        },
+      }
 
       prompt = bot.bot_prompt_with_topic_context(second_post)
 
+      req_opts = bot.reply_params.merge({ functions: bot.available_functions, stream: true })
+
       OpenAiCompletionsInferenceStubs.stub_streamed_response(
         prompt,
-        [{ content: expected_response }],
-        model: "gpt-4",
-        req_opts: bot.reply_params.merge(stream: true),
+        [expected_response],
+        model: bot.model_for,
+        req_opts: req_opts,
       )
 
-      prompt << { role: "assistant", content: "!search test search" }
-      prompt << { role: "user", content: "results: No results found" }
-      prompt << { role: "assistant", content: "!search test2 search" }
-      prompt << { role: "user", content: "results: No results found" }
+      prompt << { role: "function", content: "[]", name: "search" }
 
       OpenAiCompletionsInferenceStubs.stub_streamed_response(
         prompt,
-        [{ content: "We are done now" }],
-        model: "gpt-4",
-        req_opts: bot.reply_params.merge(stream: true),
+        [content: "I found nothing, sorry"],
+        model: bot.model_for,
+        req_opts: req_opts,
       )
 
       bot.reply_to(second_post)
@@ -75,10 +78,9 @@ RSpec.describe DiscourseAi::AiBot::Bot do
       expect(last.raw).to include("<details>")
       expect(last.raw).to include("<summary>Search</summary>")
       expect(last.raw).not_to include("translation missing")
-      expect(last.raw).to include("ok, searching...")
-      expect(last.raw).to include("We are done now")
+      expect(last.raw).to include("I found nothing")
 
-      expect(last.post_custom_prompt.custom_prompt.to_s).to include("We are done now")
+      expect(last.post_custom_prompt.custom_prompt.to_s).to include("I found nothing")
     end
   end
 
@@ -89,9 +91,9 @@ RSpec.describe DiscourseAi::AiBot::Bot do
 
     it "updates the title using bot suggestions" do
       OpenAiCompletionsInferenceStubs.stub_response(
-        [bot.title_prompt(second_post)],
+        bot.title_prompt(second_post),
         expected_response,
-        model: "gpt-4",
+        model: bot.model_for,
         req_opts: {
           temperature: 0.7,
           top_p: 0.9,

spec/lib/modules/ai_bot/commands/categories_command_spec.rb

@@ -8,8 +8,8 @@ RSpec.describe DiscourseAi::AiBot::Commands::CategoriesCommand do
       Fabricate(:category, name: "america", posts_year: 999)
 
       info = DiscourseAi::AiBot::Commands::CategoriesCommand.new(nil, nil).process(nil)
-      expect(info).to include("america")
-      expect(info).to include("999")
+      expect(info.to_s).to include("america")
+      expect(info.to_s).to include("999")
     end
   end
 end

spec/lib/modules/ai_bot/commands/command_spec.rb

@@ -14,8 +14,9 @@ RSpec.describe DiscourseAi::AiBot::Commands::Command do
       formatted =
         command.format_results(rows, column_names) { |row| ["row ¦ 1", row + 1, "a|b,\nc"] }
 
-      expect(formatted.split("\n").length).to eq(6)
-      expect(formatted).to include("a|b, c")
+      expect(formatted[:column_names].length).to eq(3)
+      expect(formatted[:rows].length).to eq(5)
+      expect(formatted.to_s).to include("a|b,\\nc")
     end
 
     it "can also generate results by returning hash per row" do

spec/lib/modules/ai_bot/commands/google_command_spec.rb

@@ -33,7 +33,7 @@ RSpec.describe DiscourseAi::AiBot::Commands::GoogleCommand do
       ).to_return(status: 200, body: json_text, headers: {})
 
       google = described_class.new(bot_user, post)
-      info = google.process("some search term")
+      info = google.process({ query: "some search term" }.to_json).to_json
 
       expect(google.description_args[:count]).to eq(1)
       expect(info).to include("title1")

spec/lib/modules/ai_bot/commands/search_command_spec.rb

@@ -13,8 +13,19 @@ RSpec.describe DiscourseAi::AiBot::Commands::SearchCommand do
       post1 = Fabricate(:post)
       search = described_class.new(bot_user, post1)
 
-      results = search.process("order:fake ABDDCDCEDGDG")
-      expect(results).to eq("No results found")
+      results = search.process({ query: "order:fake ABDDCDCEDGDG" }.to_json)
+      expect(results).to eq([])
+    end
+
+    it "supports subfolder properly" do
+      Discourse.stubs(:base_path).returns("/subfolder")
+
+      post1 = Fabricate(:post)
+
+      search = described_class.new(bot_user, post1)
+
+      results = search.process({ limit: 1, user: post1.user.username }.to_json)
+      expect(results[:rows].to_s).to include("/subfolder" + post1.url)
     end
 
     it "can handle limits" do
@@ -25,14 +36,14 @@ RSpec.describe DiscourseAi::AiBot::Commands::SearchCommand do
       # search has no built in support for limit: so handle it from the outside
       search = described_class.new(bot_user, post1)
 
-      results = search.process("@#{post1.user.username} limit:2")
+      results = search.process({ limit: 2, user: post1.user.username }.to_json)
 
-      # title + 2 rows
-      expect(results.split("\n").length).to eq(3)
+      expect(results[:column_names].length).to eq(4)
+      expect(results[:rows].length).to eq(2)
 
       # just searching for everything
-      results = search.process("order:latest_topic")
-      expect(results.split("\n").length).to be > 1
+      results = search.process({ order: "latest_topic" }.to_json)
+      expect(results[:rows].length).to be > 1
     end
   end
 end

spec/lib/modules/ai_bot/commands/tags_command_spec.rb

@@ -12,8 +12,8 @@ RSpec.describe DiscourseAi::AiBot::Commands::TagsCommand do
 
       info = DiscourseAi::AiBot::Commands::TagsCommand.new(nil, nil).process(nil)
 
-      expect(info).to include("america")
-      expect(info).not_to include("not_here")
+      expect(info.to_s).to include("america")
+      expect(info.to_s).not_to include("not_here")
     end
   end
 end

spec/lib/modules/ai_bot/jobs/regular/create_ai_reply_spec.rb

@@ -19,6 +19,7 @@ RSpec.describe Jobs::CreateAiReply do
 
       before do
         bot_user = User.find(DiscourseAi::AiBot::EntryPoint::GPT3_5_TURBO_ID)
+        bot = DiscourseAi::AiBot::Bot.as(bot_user)
 
         # time needs to be frozen so time in prompt does not drift
         freeze_time
@@ -26,10 +27,12 @@ RSpec.describe Jobs::CreateAiReply do
         OpenAiCompletionsInferenceStubs.stub_streamed_response(
           DiscourseAi::AiBot::OpenAiBot.new(bot_user).bot_prompt_with_topic_context(post),
           deltas,
+          model: bot.model_for,
           req_opts: {
             temperature: 0.4,
             top_p: 0.9,
             max_tokens: 1500,
+            functions: bot.available_functions,
             stream: true,
           },
         )
@@ -66,7 +69,7 @@ RSpec.describe Jobs::CreateAiReply do
     end
 
     context "when chatting with Claude from Anthropic" do
-      let(:claude_response) { "Assistant: #{expected_response}" }
+      let(:claude_response) { "#{expected_response}" }
       let(:deltas) { claude_response.split(" ").map { |w| "#{w} " } }
 
       before do

spec/lib/modules/ai_bot/open_ai_bot_spec.rb

@@ -33,8 +33,6 @@ RSpec.describe DiscourseAi::AiBot::OpenAiBot do
       it "trims the prompt" do
         prompt_messages = subject.bot_prompt_with_topic_context(post_1)
 
-        expect(prompt_messages[-2][:role]).to eq("assistant")
-        expect(prompt_messages[-1][:role]).to eq("user")
         # trimming is tricky... it needs to account for system message as
         # well... just make sure we trim for now
         expect(prompt_messages[-1][:content].length).to be < post_1.raw.length

spec/shared/inference/openai_completions_spec.rb

@@ -6,6 +6,98 @@ require_relative "../../support/openai_completions_inference_stubs"
 describe DiscourseAi::Inference::OpenAiCompletions do
   before { SiteSetting.ai_openai_api_key = "abc-123" }
 
+  it "supports function calling" do
+    prompt = [role: "system", content: "you are weatherbot"]
+    prompt << { role: "user", content: "what is the weather in sydney?" }
+
+    functions = []
+
+    function =
+      DiscourseAi::Inference::OpenAiCompletions::Function.new(
+        name: "get_weather",
+        description: "Get the weather in a city",
+      )
+
+    function.add_parameter(
+      name: "location",
+      type: "string",
+      description: "the city name",
+      required: true,
+    )
+
+    function.add_parameter(
+      name: "unit",
+      type: "string",
+      description: "the unit of measurement celcius c or fahrenheit f",
+      enum: %w[c f],
+      required: true,
+    )
+
+    functions << function
+
+    function_calls = []
+    current_function_call = nil
+
+    deltas = [
+      { role: "assistant" },
+      { function_call: { name: "get_weather", arguments: "" } },
+      { function_call: { arguments: "{ \"location\": " } },
+      { function_call: { arguments: "\"sydney\", \"unit\": \"c\" }" } },
+    ]
+
+    OpenAiCompletionsInferenceStubs.stub_streamed_response(
+      prompt,
+      deltas,
+      model: "gpt-3.5-turbo-0613",
+      req_opts: {
+        functions: functions,
+        stream: true,
+      },
+    )
+
+    DiscourseAi::Inference::OpenAiCompletions.perform!(
+      prompt,
+      "gpt-3.5-turbo-0613",
+      functions: functions,
+    ) do |json, cancel|
+      fn = json.dig(:choices, 0, :delta, :function_call)
+      if fn && fn[:name]
+        current_function_call = { name: fn[:name], arguments: +fn[:arguments].to_s.dup }
+        function_calls << current_function_call
+      elsif fn && fn[:arguments] && current_function_call
+        current_function_call[:arguments] << fn[:arguments]
+      end
+    end
+
+    expect(function_calls.length).to eq(1)
+    expect(function_calls[0][:name]).to eq("get_weather")
+    expect(JSON.parse(function_calls[0][:arguments])).to eq(
+      { "location" => "sydney", "unit" => "c" },
+    )
+
+    prompt << { role: "function", name: "get_weather", content: 22.to_json }
+
+    OpenAiCompletionsInferenceStubs.stub_response(
+      prompt,
+      "The current temperature in Sydney is 22 degrees Celsius.",
+      model: "gpt-3.5-turbo-0613",
+      req_opts: {
+        functions: functions,
+      },
+    )
+
+    result =
+      DiscourseAi::Inference::OpenAiCompletions.perform!(
+        prompt,
+        "gpt-3.5-turbo-0613",
+        functions: functions,
+      )
+
+    expect(result.dig(:choices, 0, :message, :content)).to eq(
+      "The current temperature in Sydney is 22 degrees Celsius.",
+    )
+  end
+
   it "can complete a trivial prompt" do
     response_text = "1. Serenity\\n2. Laughter\\n3. Adventure"
     prompt = [role: "user", content: "write 3 words"]