FEATURE: AI artifacts (#898)

This is a significant PR that introduces AI Artifacts functionality to the discourse-ai plugin along with several other improvements. Here are the key changes: 1. AI Artifacts System: - Adds a new `AiArtifact` model and database migration - Allows creation of web artifacts with HTML, CSS, and JavaScript content - Introduces security settings (`strict`, `lax`, `disabled`) for controlling artifact execution - Implements artifact rendering in iframes with sandbox protection - New `CreateArtifact` tool for AI to generate interactive content 2. Tool System Improvements: - Adds support for partial tool calls, allowing incremental updates during generation - Better handling of tool call states and progress tracking - Improved XML tool processing with CDATA support - Fixes for tool parameter handling and duplicate invocations 3. LLM Provider Updates: - Updates for Anthropic Claude models with correct token limits - Adds support for native/XML tool modes in Gemini integration - Adds new model configurations including Llama 3.1 models - Improvements to streaming response handling 4. UI Enhancements: - New artifact viewer component with expand/collapse functionality - Security controls for artifact execution (click-to-run in strict mode) - Improved dialog and response handling - Better error management for tool execution 5. Security Improvements: - Sandbox controls for artifact execution - Public/private artifact sharing controls - Security settings to control artifact behavior - CSP and frame-options handling for artifacts 6. Technical Improvements: - Better post streaming implementation - Improved error handling in completions - Better memory management for partial tool calls - Enhanced testing coverage 7. Configuration: - New site settings for artifact security - Extended LLM model configurations - Additional tool configuration options This PR significantly enhances the plugin's capabilities for generating and displaying interactive content while maintaining security and providing flexible configuration options for administrators.
2024-11-19 09:22:39 +11:00 · 2024-11-19 09:22:39 +11:00 · 0d7f353284
parent 4fb686a548
commit 0d7f353284
46 changed files with 1707 additions and 179 deletions
--- a/app/controllers/discourse_ai/ai_bot/artifacts_controller.rb
+++ b/app/controllers/discourse_ai/ai_bot/artifacts_controller.rb
@ -0,0 +1,59 @@
+# frozen_string_literal: true
+
+module DiscourseAi
+  module AiBot
+    class ArtifactsController < ApplicationController
+      requires_plugin DiscourseAi::PLUGIN_NAME
+      before_action :require_site_settings!
+
+      skip_before_action :preload_json, :check_xhr, only: %i[show]
+
+      def show
+        artifact = AiArtifact.find(params[:id])
+
+        post = Post.find_by(id: artifact.post_id)
+        if artifact.metadata&.dig("public")
+          # no guardian needed
+        else
+          raise Discourse::NotFound if !post&.topic&.private_message?
+          raise Discourse::NotFound if !guardian.can_see?(post)
+        end
+
+        # Prepare the HTML document
+        html = <<~HTML
+          <!DOCTYPE html>
+          <html>
+            <head>
+              <meta charset="UTF-8">
+              <title>#{ERB::Util.html_escape(artifact.name)}</title>
+              <style>
+                #{artifact.css}
+              </style>
+            </head>
+            <body>
+              #{artifact.html}
+              <script>
+                #{artifact.js}
+              </script>
+            </body>
+          </html>
+        HTML
+
+        response.headers.delete("X-Frame-Options")
+        response.headers.delete("Content-Security-Policy")
+
+        # Render the content
+        render html: html.html_safe, layout: false, content_type: "text/html"
+      end
+
+      private
+
+      def require_site_settings!
+        if !SiteSetting.discourse_ai_enabled ||
+             !SiteSetting.ai_artifact_security.in?(%w[lax strict])
+          raise Discourse::NotFound
+        end
+      end
+    end
+  end
+end
--- a/app/models/ai_artifact.rb
+++ b/app/models/ai_artifact.rb
@ -0,0 +1,49 @@
+# frozen_string_literal: true
+
+class AiArtifact < ActiveRecord::Base
+  belongs_to :user
+  belongs_to :post
+  validates :html, length: { maximum: 65_535 }
+  validates :css, length: { maximum: 65_535 }
+  validates :js, length: { maximum: 65_535 }
+
+  def self.iframe_for(id)
+    <<~HTML
+      <iframe sandbox="allow-scripts allow-forms" height="600px" src='#{url(id)}' frameborder="0" width="100%"></iframe>
+    HTML
+  end
+
+  def self.url(id)
+    Discourse.base_url + "/discourse-ai/ai-bot/artifacts/#{id}"
+  end
+
+  def self.share_publicly(id:, post:)
+    artifact = AiArtifact.find_by(id: id)
+    artifact.update!(metadata: { public: true }) if artifact&.post&.topic&.id == post.topic.id
+  end
+
+  def self.unshare_publicly(id:)
+    artifact = AiArtifact.find_by(id: id)
+    artifact&.update!(metadata: { public: false })
+  end
+
+  def url
+    self.class.url(id)
+  end
+end
+
+# == Schema Information
+#
+# Table name: ai_artifacts
+#
+#  id         :bigint           not null, primary key
+#  user_id    :integer          not null
+#  post_id    :integer          not null
+#  name       :string(255)      not null
+#  html       :string(65535)
+#  css        :string(65535)
+#  js         :string(65535)
+#  metadata   :jsonb
+#  created_at :datetime         not null
+#  updated_at :datetime         not null
+#
--- a/app/models/llm_model.rb
+++ b/app/models/llm_model.rb
@ -26,6 +26,13 @@ class LlmModel < ActiveRecord::Base
      },
      open_ai: {
        organization: :text,
+        disable_native_tools: :checkbox,
+      },
+      google: {
+        disable_native_tools: :checkbox,
+      },
+      azure: {
+        disable_native_tools: :checkbox,
      },
      hugging_face: {
        disable_system_prompt: :checkbox,
--- a/app/models/shared_ai_conversation.rb
+++ b/app/models/shared_ai_conversation.rb
@ -34,6 +34,12 @@ class SharedAiConversation < ActiveRecord::Base

  def self.destroy_conversation(conversation)
    conversation.destroy
+
+    maybe_topic = conversation.target
+    if maybe_topic.is_a?(Topic)
+      AiArtifact.where(post: maybe_topic.posts).update_all(metadata: { public: false })
+    end
+
    ::Jobs.enqueue(
      :shared_conversation_adjust_upload_security,
      target_id: conversation.target_id,
@ -165,7 +171,7 @@ class SharedAiConversation < ActiveRecord::Base
            id: post.id,
            user_id: post.user_id,
            created_at: post.created_at,
-            cooked: post.cooked,
+            cooked: cook_artifacts(post),
          }

          mapped[:persona] = persona if ai_bot_participant&.id == post.user_id
@ -175,6 +181,24 @@ class SharedAiConversation < ActiveRecord::Base
    }
  end

+  def self.cook_artifacts(post)
+    html = post.cooked
+    return html if !%w[lax strict].include?(SiteSetting.ai_artifact_security)
+
+    doc = Nokogiri::HTML5.fragment(html)
+    doc
+      .css("div.ai-artifact")
+      .each do |node|
+        id = node["data-ai-artifact-id"].to_i
+        if id > 0
+          AiArtifact.share_publicly(id: id, post: post)
+          node.replace(AiArtifact.iframe_for(id))
+        end
+      end
+
+    doc.to_s
+  end
+
  private

  def populate_user_info!(posts)
--- a/assets/javascripts/discourse/components/ai-artifact.gjs
+++ b/assets/javascripts/discourse/components/ai-artifact.gjs
@ -0,0 +1,122 @@
+import Component from "@glimmer/component";
+import { tracked } from "@glimmer/tracking";
+import { on } from "@ember/modifier";
+import { action } from "@ember/object";
+import { inject as service } from "@ember/service";
+import DButton from "discourse/components/d-button";
+import htmlClass from "discourse/helpers/html-class";
+import getURL from "discourse-common/lib/get-url";
+
+export default class AiArtifactComponent extends Component {
+  @service siteSettings;
+  @tracked expanded = false;
+  @tracked showingArtifact = false;
+
+  constructor() {
+    super(...arguments);
+    this.keydownHandler = this.handleKeydown.bind(this);
+  }
+
+  willDestroy() {
+    super.willDestroy(...arguments);
+    window.removeEventListener("keydown", this.keydownHandler);
+  }
+
+  @action
+  handleKeydown(event) {
+    if (event.key === "Escape" || event.key === "Esc") {
+      this.expanded = false;
+    }
+  }
+
+  get requireClickToRun() {
+    if (this.showingArtifact) {
+      return false;
+    }
+    return this.siteSettings.ai_artifact_security === "strict";
+  }
+
+  get artifactUrl() {
+    return getURL(`/discourse-ai/ai-bot/artifacts/${this.args.artifactId}`);
+  }
+
+  @action
+  showArtifact() {
+    this.showingArtifact = true;
+  }
+
+  @action
+  toggleView() {
+    this.expanded = !this.expanded;
+    if (this.expanded) {
+      window.addEventListener("keydown", this.keydownHandler);
+    } else {
+      window.removeEventListener("keydown", this.keydownHandler);
+    }
+  }
+
+  get wrapperClasses() {
+    return `ai-artifact__wrapper ${
+      this.expanded ? "ai-artifact__expanded" : ""
+    }`;
+  }
+
+  @action
+  artifactPanelHover() {
+    // retrrigger animation
+    const panel = document.querySelector(".ai-artifact__panel");
+    panel.style.animation = "none"; // Stop the animation
+    setTimeout(() => {
+      panel.style.animation = ""; // Re-trigger the animation by removing the none style
+    }, 0);
+  }
+
+  <template>
+    {{#if this.expanded}}
+      {{htmlClass "ai-artifact-expanded"}}
+    {{/if}}
+    <div class={{this.wrapperClasses}}>
+      <div
+        class="ai-artifact__panel--wrapper"
+        {{on "mouseleave" this.artifactPanelHover}}
+      >
+        <div class="ai-artifact__panel">
+          <DButton
+            class="btn-flat btn-icon-text"
+            @icon="discourse-compress"
+            @label="discourse_ai.ai_artifact.collapse_view_label"
+            @action={{this.toggleView}}
+          />
+        </div>
+      </div>
+      {{#if this.requireClickToRun}}
+        <div class="ai-artifact__click-to-run">
+          <DButton
+            class="btn btn-primary"
+            @icon="play"
+            @label="discourse_ai.ai_artifact.click_to_run_label"
+            @action={{this.showArtifact}}
+          />
+        </div>
+      {{else}}
+        <iframe
+          title="AI Artifact"
+          src={{this.artifactUrl}}
+          width="100%"
+          frameborder="0"
+          sandbox="allow-scripts allow-forms"
+        ></iframe>
+      {{/if}}
+      {{#unless this.requireClickToRun}}
+        <div class="ai-artifact__footer">
+          <DButton
+            class="btn-flat btn-icon-text ai-artifact__expand-button"
+            @icon="discourse-expand"
+            @label="discourse_ai.ai_artifact.expand_view_label"
+            @action={{this.toggleView}}
+          />
+        </div>
+      {{/unless}}
+    </div>
+  </template>
+}
--- a/assets/javascripts/initializers/ai-artifacts.gjs
+++ b/assets/javascripts/initializers/ai-artifacts.gjs
@ -0,0 +1,35 @@
+import { withPluginApi } from "discourse/lib/plugin-api";
+import AiArtifact from "../discourse/components/ai-artifact";
+
+function initializeAiArtifacts(api) {
+  api.decorateCookedElement(
+    (element, helper) => {
+      if (!helper.renderGlimmer) {
+        return;
+      }
+
+      [...element.querySelectorAll("div.ai-artifact")].forEach(
+        (artifactElement) => {
+          const artifactId = artifactElement.getAttribute(
+            "data-ai-artifact-id"
+          );
+
+          helper.renderGlimmer(artifactElement, <template>
+            <AiArtifact @artifactId={{artifactId}} />
+          </template>);
+        }
+      );
+    },
+    {
+      id: "ai-artifact",
+      onlyStream: true,
+    }
+  );
+}
+
+export default {
+  name: "ai-artifact",
+  initialize() {
+    withPluginApi("0.8.7", initializeAiArtifacts);
+  },
+};
--- a/assets/javascripts/lib/discourse-markdown/ai-tags.js
+++ b/assets/javascripts/lib/discourse-markdown/ai-tags.js
@ -1,3 +1,4 @@
 export function setup(helper) {
  helper.allowList(["details[class=ai-quote]"]);
+  helper.allowList(["div[class=ai-artifact]", "div[data-ai-artifact-id]"]);
 }
--- a/assets/stylesheets/modules/ai-bot/common/ai-artifact.scss
+++ b/assets/stylesheets/modules/ai-bot/common/ai-artifact.scss
@ -0,0 +1,106 @@
+.ai-artifact__wrapper {
+  iframe {
+    width: 100%;
+    height: calc(100% - 2em);
+  }
+  height: 500px;
+  padding-bottom: 2em;
+}
+
+.ai-artifact__click-to-run {
+  display: flex;
+  justify-content: center;
+  align-items: center;
+  height: 100%;
+}
+
+.ai-artifact__panel {
+  display: none;
+}
+
+html.ai-artifact-expanded {
+  overflow: hidden;
+}
+
+.ai-artifact__footer {
+  display: flex;
+  justify-content: space-between;
+  align-items: center;
+  .ai-artifact__expand-button {
+    margin-left: auto;
+  }
+}
+
+.ai-artifact__expanded {
+  .ai-artifact__footer {
+    display: none;
+  }
+
+  .ai-artifact__panel--wrapper {
+    display: block;
+    position: fixed;
+    top: 0;
+    left: 0;
+    right: 0;
+    height: 4em;
+    z-index: 1000000;
+    &:hover {
+      .ai-artifact__panel {
+        transform: translateY(0) !important;
+        animation: none;
+      }
+    }
+  }
+
+  .ai-artifact__panel {
+    display: block;
+    position: fixed;
+    top: 0;
+    left: 0;
+    right: 0;
+    height: 2em;
+    transition: transform 0.5s ease-in-out;
+    animation: slideUp 0.5s 3s forwards;
+    background-color: var(--secondary-low);
+    opacity: 0.9;
+    transform: translateY(0);
+    button {
+      width: 100%;
+      text-align: left;
+      box-sizing: border-box;
+      justify-content: flex-start;
+      color: var(--secondary-very-high);
+      &:hover {
+        color: var(--secondary-very-high);
+        .d-icon {
+          color: var(--secondary-high);
+        }
+        //color: var(--secondary-vary-low);
+      }
+    }
+  }
+  @keyframes slideUp {
+    to {
+      transform: translateY(-100%);
+    }
+  }
+
+  iframe {
+    position: fixed;
+    top: 0;
+    height: 100%;
+    max-height: 100%;
+    left: 0;
+    right: 0;
+    bottom: 0;
+    z-index: z("fullscreen");
+  }
+
+  position: fixed;
+  top: 0;
+  left: 0;
+  height: 100%;
+  width: 100%;
+  z-index: z("fullscreen");
+  background-color: var(--secondary);
+}
--- a/config/locales/client.en.yml
+++ b/config/locales/client.en.yml
@ -165,7 +165,7 @@ en:
        saved: "Persona saved"
        enabled: "Enabled?"
        tools: "Enabled tools"
-        forced_tools: "Forced fools"
+        forced_tools: "Forced tools"
        allowed_groups: "Allowed groups"
        confirm_delete: "Are you sure you want to delete this persona?"
        new: "New persona"
@ -279,14 +279,17 @@ en:
        model_description:
          none: "General settings that work for most language models"
          anthropic-claude-3-5-sonnet: "Anthropic's most intelligent model"
+          anthropic-claude-3-5-haiku: "Fast and cost-effective"
          anthropic-claude-3-opus: "Excels at writing and complex tasks"
-          anthropic-claude-3-sonnet: "Balance of speed and intelligence"
-          anthropic-claude-3-haiku: "Fast and cost-effective"
          google-gemini-1-5-pro: "Mid-sized multimodal model capable of a wide range of tasks"
          google-gemini-1-5-flash: "Lightweight, fast, and cost-efficient with multimodal reasoning"
          open_ai-gpt-4-turbo: "Previous generation high-intelligence model"
          open_ai-gpt-4o: "High intelligence model for complex, multi-step tasks"
          open_ai-gpt-4o-mini: "Affordable and fast small model for lightweight tasks"
+          open_ai-o1-mini: "Cost-efficient reasoning model"
+          open_ai-o1-preview: "Open AI's most capabale reasoning model"
+          samba_nova-Meta-Llama-3-1-8B-Instruct: "Efficient lightweight multilingual model"
+          samba_nova-Meta-Llama-3-1-70B-Instruct": "Powerful multipurpose model"

        configured:
          title: "Configured LLMs"
@ -399,6 +402,11 @@ en:
        quick_search:
          suffix: "in all topics and posts with AI"

+      ai_artifact:
+        expand_view_label: "Expand view"
+        collapse_view_label: "Exit Fullscreen (ESC)"
+        click_to_run_label: "Run Artifact"
+
      ai_bot:
        pm_warning: "AI chatbot messages are monitored regularly by moderators."
        cancel_streaming: "Stop reply"
--- a/config/locales/server.en.yml
+++ b/config/locales/server.en.yml
@ -17,6 +17,7 @@ en:
        description: "Periodic report based on a large language model"
  site_settings:
    discourse_ai_enabled: "Enable the discourse AI plugin."
+    ai_artifact_security: "The AI artifact system generates IFRAMEs with runnable code. Strict mode disables sharing and forces an extra click to run code. Lax mode allows sharing of artifacts and runs code directly. Disabled mode disables the artifact system."
    ai_toxicity_enabled: "Enable the toxicity module."
    ai_toxicity_inference_service_api_endpoint: "URL where the API is running for the toxicity module"
    ai_toxicity_inference_service_api_key: "API key for the toxicity API"
@ -199,6 +200,9 @@ en:
        discourse_helper:
          name: "Discourse Helper"
          description: "AI Bot specialized in helping with Discourse related tasks"
+        web_artifact_creator:
+          name: "Web Artifact Creator"
+          description: "AI Bot specialized in creating interactive web artifacts"
      topic_not_found: "Summary unavailable, topic not found!"
      summarizing: "Summarizing topic"
      searching: "Searching for: '%{query}'"
@ -222,6 +226,7 @@ en:
            name: "Base Search Query"
            description: "Base query to use when searching. Example: '#urgent' will prepend '#urgent' to the search query and only include topics with the urgent category or tag."
      tool_summary:
+        create_artifact: "Create web artifact"
        web_browser: "Browse Web"
        github_search_files: "GitHub search files"
        github_search_code: "GitHub code search"
@ -243,6 +248,7 @@ en:
        search_meta_discourse: "Search Meta Discourse"
        javascript_evaluator: "Evaluate JavaScript"
      tool_help:
+        create_artifact: "Create a web artifact using the AI Bot"
        web_browser: "Browse web page using the AI Bot"
        github_search_code: "Search for code in a GitHub repository"
        github_search_files: "Search for files in a GitHub repository"
@ -264,6 +270,7 @@ en:
        search_meta_discourse: "Search Meta Discourse"
        javascript_evaluator: "Evaluate JavaScript"
      tool_description:
+        create_artifact: "Created a web artifact using the AI Bot"
        web_browser: "Reading <a href='%{url}'>%{url}</a>"
        github_search_files: "Searched for '%{keywords}' in %{repo}/%{branch}"
        github_search_code: "Searched for '%{query}' in %{repo}"
--- a/config/routes.rb
+++ b/config/routes.rb
@ -33,6 +33,10 @@ DiscourseAi::Engine.routes.draw do
    get "/preview/:topic_id" => "shared_ai_conversations#preview"
  end

+  scope module: :ai_bot, path: "/ai-bot/artifacts" do
+    get "/:id" => "artifacts#show"
+  end
+
  scope module: :summarization, path: "/summarization", defaults: { format: :json } do
    get "/t/:topic_id" => "summary#show", :constraints => { topic_id: /\d+/ }
    get "/channels/:channel_id" => "chat_summary#show"
--- a/config/settings.yml
+++ b/config/settings.yml
@ -2,7 +2,14 @@ discourse_ai:
  discourse_ai_enabled:
    default: true
    client: true
-
+  ai_artifact_security:
+    client: true
+    type: enum
+    default: "strict"
+    choices:
+      - "disabled"
+      - "lax"
+      - "strict"
  ai_toxicity_enabled:
    default: false
    client: true
--- a/db/fixtures/ai_bot/603_bot_ai_personas.rb
+++ b/db/fixtures/ai_bot/603_bot_ai_personas.rb
@ -5,7 +5,12 @@ DiscourseAi::AiBot::Personas::Persona.system_personas.each do |persona_class, id
  if !persona
    persona = AiPersona.new
    persona.id = id
-    persona.allowed_group_ids = [Group::AUTO_GROUPS[:trust_level_0]]
+    if persona_class == DiscourseAi::AiBot::Personas::WebArtifactCreator
+      # this is somewhat sensitive, so we default it to staff
+      persona.allowed_group_ids = [Group::AUTO_GROUPS[:staff]]
+    else
+      persona.allowed_group_ids = [Group::AUTO_GROUPS[:trust_level_0]]
+    end
    persona.enabled = true
    persona.priority = true if persona_class == DiscourseAi::AiBot::Personas::General
  end
--- a/db/migrate/20241104053017_add_ai_artifacts.rb
+++ b/db/migrate/20241104053017_add_ai_artifacts.rb
@ -0,0 +1,16 @@
+# frozen_string_literal: true
+class AddAiArtifacts < ActiveRecord::Migration[7.1]
+  def change
+    create_table :ai_artifacts do |t|
+      t.integer :user_id, null: false
+      t.integer :post_id, null: false
+      t.string :name, null: false, limit: 255
+      t.string :html, limit: 65_535 # ~64KB limit
+      t.string :css, limit: 65_535 # ~64KB limit
+      t.string :js, limit: 65_535 # ~64KB limit
+      t.jsonb :metadata # For any additional properties
+
+      t.timestamps
+    end
+  end
+end
--- a/lib/ai_bot/bot.rb
+++ b/lib/ai_bot/bot.rb
@ -106,12 +106,39 @@ module DiscourseAi
          tool_found = false
          force_tool_if_needed(prompt, context)

+          tool_halted = false
+
+          allow_partial_tool_calls = persona.allow_partial_tool_calls?
+          existing_tools = Set.new
+
          result =
-            llm.generate(prompt, feature_name: "bot", **llm_kwargs) do |partial, cancel|
-              tool = persona.find_tool(partial, bot_user: user, llm: llm, context: context)
+            llm.generate(
+              prompt,
+              feature_name: "bot",
+              partial_tool_calls: allow_partial_tool_calls,
+              **llm_kwargs,
+            ) do |partial, cancel|
+              tool =
+                persona.find_tool(
+                  partial,
+                  bot_user: user,
+                  llm: llm,
+                  context: context,
+                  existing_tools: existing_tools,
+                )
              tool = nil if tools_ran >= MAX_TOOLS

              if tool.present?
+                existing_tools << tool
+                tool_call = partial
+                if tool_call.partial?
+                  if tool.class.allow_partial_tool_calls?
+                    tool.partial_invoke
+                    update_blk.call("", cancel, tool.custom_raw, :partial_tool)
+                  end
+                  next
+                end
+
                tool_found = true
                # a bit hacky, but extra newlines do no harm
                if needs_newlines
@ -122,7 +149,10 @@ module DiscourseAi
                process_tool(tool, raw_context, llm, cancel, update_blk, prompt, context)
                tools_ran += 1
                ongoing_chain &&= tool.chain_next_response?
+
+                tool_halted = true if !tool.chain_next_response?
              else
+                next if tool_halted
                needs_newlines = true
                if partial.is_a?(DiscourseAi::Completions::ToolCall)
                  Rails.logger.warn("DiscourseAi: Tool not found: #{partial.name}")
@ -185,7 +215,7 @@ module DiscourseAi
      end

      def invoke_tool(tool, llm, cancel, context, &update_blk)
-        show_placeholder = !context[:skip_tool_details]
+        show_placeholder = !context[:skip_tool_details] && !tool.class.allow_partial_tool_calls?

        update_blk.call("", cancel, build_placeholder(tool.summary, "")) if show_placeholder

--- a/lib/ai_bot/personas/persona.rb
+++ b/lib/ai_bot/personas/persona.rb
@ -44,6 +44,7 @@ module DiscourseAi
              Personas::DallE3 => -7,
              Personas::DiscourseHelper => -8,
              Personas::GithubHelper => -9,
+              Personas::WebArtifactCreator => -10,
            }
          end

@ -98,6 +99,7 @@ module DiscourseAi
              Tools::JavascriptEvaluator,
            ]

+            tools << Tools::CreateArtifact if SiteSetting.ai_artifact_security.in?(%w[lax strict])
            tools << Tools::GithubSearchCode if SiteSetting.ai_bot_github_access_token.present?

            tools << Tools::ListTags if SiteSetting.tagging_enabled
@ -199,14 +201,24 @@ module DiscourseAi
          prompt
        end

-        def find_tool(partial, bot_user:, llm:, context:)
+        def find_tool(partial, bot_user:, llm:, context:, existing_tools: [])
          return nil if !partial.is_a?(DiscourseAi::Completions::ToolCall)
-          tool_instance(partial, bot_user: bot_user, llm: llm, context: context)
+          tool_instance(
+            partial,
+            bot_user: bot_user,
+            llm: llm,
+            context: context,
+            existing_tools: existing_tools,
+          )
+        end
+
+        def allow_partial_tool_calls?
+          available_tools.any? { |tool| tool.allow_partial_tool_calls? }
        end

        protected

-        def tool_instance(tool_call, bot_user:, llm:, context:)
+        def tool_instance(tool_call, bot_user:, llm:, context:, existing_tools:)
          function_id = tool_call.id
          function_name = tool_call.name
          return nil if function_name.nil?
@ -240,14 +252,22 @@ module DiscourseAi
            arguments[name.to_sym] = value if value
          end

-          tool_klass.new(
-            arguments,
-            tool_call_id: function_id || function_name,
-            persona_options: options[tool_klass].to_h,
-            bot_user: bot_user,
-            llm: llm,
-            context: context,
-          )
+          tool_instance =
+            existing_tools.find { |t| t.name == function_name && t.tool_call_id == function_id }
+
+          if tool_instance
+            tool_instance.parameters = arguments
+            tool_instance
+          else
+            tool_klass.new(
+              arguments,
+              tool_call_id: function_id || function_name,
+              persona_options: options[tool_klass].to_h,
+              bot_user: bot_user,
+              llm: llm,
+              context: context,
+            )
+          end
        end

        def strip_quotes(value)
--- a/lib/ai_bot/personas/web_artifact_creator.rb
+++ b/lib/ai_bot/personas/web_artifact_creator.rb
@ -0,0 +1,56 @@
+#frozen_string_literal: true
+
+module DiscourseAi
+  module AiBot
+    module Personas
+      class WebArtifactCreator < Persona
+        def tools
+          [Tools::CreateArtifact]
+        end
+
+        def required_tools
+          [Tools::CreateArtifact]
+        end
+
+        def system_prompt
+          <<~PROMPT
+            You are the Web Creator, an AI assistant specializing in building interactive web components. You create engaging and functional web experiences using HTML, CSS, and JavaScript. You live in a Discourse PM and communicate using Markdown.
+
+            Core Principles:
+            - Create delightful, interactive experiences
+            - Focus on visual appeal and smooth animations
+            - Write clean, efficient code
+            - Build progressively (HTML structure → CSS styling → JavaScript interactivity)
+            - Keep components focused and purposeful
+
+            When creating:
+            1. Understand the desired user experience
+            2. Break down complex interactions into simple components
+            3. Use semantic HTML for strong foundations
+            4. Style thoughtfully with CSS
+            5. Add JavaScript for rich interactivity
+            6. Consider responsive design
+
+            Best Practices:
+            - Leverage native HTML elements for better functionality
+            - Use CSS transforms and transitions for smooth animations
+            - Keep JavaScript modular and event-driven
+            - Make content responsive and adaptive
+            - Create self-contained components
+
+            When responding:
+            1. Ask clarifying questions if the request is ambiguous
+            2. Briefly explain your approach
+            3. Build features iteratively
+            4. Describe the interactive elements
+            5. Test your solution conceptually
+
+            Your goal is to transform ideas into engaging web experiences. Be creative and practical, focusing on making interfaces that are both beautiful and functional.
+
+            Remember: Great components combine structure (HTML), presentation (CSS), and behavior (JavaScript) to create memorable user experiences.
+          PROMPT
+        end
+      end
+    end
+  end
+end
--- a/lib/ai_bot/playground.rb
+++ b/lib/ai_bot/playground.rb
@ -399,7 +399,7 @@ module DiscourseAi
        PostCustomPrompt.none

        reply = +""
-        start = Time.now
+        post_streamer = nil

        post_type =
          post.post_type == Post.types[:whisper] ? Post.types[:whisper] : Post.types[:regular]
@ -448,35 +448,35 @@ module DiscourseAi

        context[:skip_tool_details] ||= !bot.persona.class.tool_details

+        post_streamer = PostStreamer.new(delay: Rails.env.test? ? 0 : 0.5) if stream_reply
+
        new_custom_prompts =
          bot.reply(context) do |partial, cancel, placeholder, type|
            reply << partial
            raw = reply.dup
            raw << "\n\n" << placeholder if placeholder.present?

-            blk.call(partial) if blk && type != :tool_details
+            blk.call(partial) if blk && type != :tool_details && type != :partial_tool

            if stream_reply && !Discourse.redis.get(redis_stream_key)
              cancel&.call
              reply_post.update!(raw: reply, cooked: PrettyText.cook(reply))
            end

-            if stream_reply
-              # Minor hack to skip the delay during tests.
-              if placeholder.blank?
-                next if (Time.now - start < 0.5) && !Rails.env.test?
-                start = Time.now
+            if post_streamer
+              post_streamer.run_later do
+                Discourse.redis.expire(redis_stream_key, 60)
+                publish_update(reply_post, { raw: raw })
              end
-
-              Discourse.redis.expire(redis_stream_key, 60)
-
-              publish_update(reply_post, { raw: raw })
            end
          end

        return if reply.blank?

        if stream_reply
+          post_streamer.finish
+          post_streamer = nil
+
          # land the final message prior to saving so we don't clash
          reply_post.cooked = PrettyText.cook(reply)
          publish_final_update(reply_post)
@ -514,6 +514,7 @@ module DiscourseAi

        reply_post
      ensure
+        post_streamer&.finish(skip_callback: true)
        publish_final_update(reply_post) if stream_reply
        if reply_post && post.post_number == 1 && post.topic.private_message?
          title_playground(reply_post)
--- a/lib/ai_bot/post_streamer.rb
+++ b/lib/ai_bot/post_streamer.rb
@ -0,0 +1,58 @@
+# frozen_string_literal: true
+
+module DiscourseAi
+  module AiBot
+    class PostStreamer
+      def initialize(delay: 0.5)
+        @mutex = Mutex.new
+        @callback = nil
+        @delay = delay
+        @done = false
+      end
+
+      def run_later(&callback)
+        @mutex.synchronize { @callback = callback }
+        ensure_worker!
+      end
+
+      def finish(skip_callback: false)
+        @mutex.synchronize do
+          @callback&.call if skip_callback
+          @callback = nil
+          @done = true
+        end
+
+        begin
+          @worker_thread&.wakeup
+        rescue StandardError
+          ThreadError
+        end
+        @worker_thread&.join
+        @worker_thread = nil
+      end
+
+      private
+
+      def run
+        while !@done
+          @mutex.synchronize do
+            callback = @callback
+            @callback = nil
+            callback&.call
+          end
+          sleep @delay
+        end
+      end
+
+      def ensure_worker!
+        return if @worker_thread
+        @mutex.synchronize do
+          return if @worker_thread
+          db = RailsMultisite::ConnectionManagement.current_db
+          @worker_thread =
+            Thread.new { RailsMultisite::ConnectionManagement.with_connection(db) { run } }
+        end
+      end
+    end
+  end
+end
--- a/lib/ai_bot/tools/create_artifact.rb
+++ b/lib/ai_bot/tools/create_artifact.rb
@ -0,0 +1,137 @@
+# frozen_string_literal: true
+
+module DiscourseAi
+  module AiBot
+    module Tools
+      class CreateArtifact < Tool
+        def self.name
+          "create_artifact"
+        end
+
+        def self.signature
+          {
+            name: "create_artifact",
+            description:
+              "Creates a web artifact with HTML, CSS, and JavaScript that can be displayed in an iframe",
+            parameters: [
+              {
+                name: "name",
+                description: "A name for the artifact (max 255 chars)",
+                type: "string",
+                required: true,
+              },
+              {
+                name: "html_body",
+                description: "The HTML content for the BODY tag (do not include the BODY tag)",
+                type: "string",
+                required: true,
+              },
+              { name: "css", description: "Optional CSS styles for the artifact", type: "string" },
+              {
+                name: "js",
+                description:
+                  "Optional
+              JavaScript code for the artifact",
+                type: "string",
+              },
+            ],
+          }
+        end
+
+        def self.allow_partial_tool_calls?
+          true
+        end
+
+        def partial_invoke
+          @selected_tab = :html_body
+          if @prev_parameters
+            @selected_tab = parameters.keys.find { |k| @prev_parameters[k] != parameters[k] }
+          end
+          update_custom_html
+          @prev_parameters = parameters.dup
+        end
+
+        def invoke
+          yield parameters[:name] || "Web Artifact"
+          # Get the current post from context
+          post = Post.find_by(id: context[:post_id])
+          return error_response("No post context found") unless post
+
+          html = parameters[:html_body].to_s
+          css = parameters[:css].to_s
+          js = parameters[:js].to_s
+
+          # Create the artifact
+          artifact =
+            AiArtifact.new(
+              user_id: bot_user.id,
+              post_id: post.id,
+              name: parameters[:name].to_s[0...255],
+              html: html,
+              css: css,
+              js: js,
+              metadata: parameters[:metadata],
+            )
+
+          if artifact.save
+            update_custom_html(artifact)
+            success_response(artifact)
+          else
+            error_response(artifact.errors.full_messages.join(", "))
+          end
+        end
+
+        def chain_next_response?
+          @chain_next_response
+        end
+
+        private
+
+        def update_custom_html(artifact = nil)
+          html = parameters[:html_body].to_s
+          css = parameters[:css].to_s
+          js = parameters[:js].to_s
+
+          artifact_div =
+            "<div class=\"ai-artifact\" data-ai-artifact-id=#{artifact.id}></div>" if artifact
+
+          content = []
+
+          content << [:html_body, "### HTML\n\n```html\n#{html}\n```"] if html.present?
+
+          content << [:css, "### CSS\n\n```css\n#{css}\n```"] if css.present?
+
+          content << [:js, "### JavaScript\n\n```javascript\n#{js}\n```"] if js.present?
+
+          content << [:preview, "### Preview\n\n#{artifact_div}"] if artifact_div
+
+          content.sort_by! { |c| c[0] === @selected_tab ? 1 : 0 } if !artifact
+
+          self.custom_raw = content.map { |c| c[1] }.join("\n\n")
+        end
+
+        def success_response(artifact)
+          @chain_next_response = false
+
+          {
+            status: "success",
+            artifact_id: artifact.id,
+            message: "Artifact created successfully and rendered to user.",
+          }
+        end
+
+        def error_response(message)
+          @chain_next_response = false
+
+          { status: "error", error: message }
+        end
+
+        def help
+          "Creates a web artifact with HTML, CSS, and JavaScript that can be displayed in an iframe. " \
+            "Requires a name and HTML content. CSS and JavaScript are optional. " \
+            "The artifact will be associated with the current post and can be displayed using an iframe."
+        end
+      end
+    end
+  end
+end
--- a/lib/ai_bot/tools/tool.rb
+++ b/lib/ai_bot/tools/tool.rb
@ -38,10 +38,14 @@ module DiscourseAi
          def custom_system_message
            nil
          end
+
+          def allow_partial_tool_calls?
+            false
+          end
        end

-        attr_accessor :custom_raw
-        attr_reader :tool_call_id, :persona_options, :bot_user, :llm, :context, :parameters
+        attr_accessor :custom_raw, :parameters
+        attr_reader :tool_call_id, :persona_options, :bot_user, :llm, :context

        def initialize(
          parameters,
--- a/lib/completions/anthropic_message_processor.rb
+++ b/lib/completions/anthropic_message_processor.rb
@ -35,6 +35,8 @@ class DiscourseAi::Completions::AnthropicMessageProcessor

    def to_tool_call
      parameters = JSON.parse(raw_json, symbolize_names: true)
+      # we dupe to avoid poisoning the original tool call
+      @tool_call = @tool_call.dup
      @tool_call.partial = false
      @tool_call.parameters = parameters
      @tool_call
--- a/lib/completions/dialects/chat_gpt.rb
+++ b/lib/completions/dialects/chat_gpt.rb
@ -44,17 +44,31 @@ module DiscourseAi
          llm_model.provider == "open_ai" && llm_model.name.include?("o1-")
        end

+        def disable_native_tools?
+          return @disable_native_tools if defined?(@disable_native_tools)
+          !!@disable_native_tools = llm_model.lookup_custom_param("disable_native_tools")
+        end
+
        private

        def tools_dialect
-          @tools_dialect ||= DiscourseAi::Completions::Dialects::OpenAiTools.new(prompt.tools)
+          if disable_native_tools?
+            super
+          else
+            @tools_dialect ||= DiscourseAi::Completions::Dialects::OpenAiTools.new(prompt.tools)
+          end
        end

        def system_msg(msg)
+          content = msg[:content]
+          if disable_native_tools? && tools_dialect.instructions.present?
+            content = content + "\n\n" + tools_dialect.instructions
+          end
+
          if is_gpt_o?
-            { role: "user", content: msg[:content] }
+            { role: "user", content: content }
          else
-            { role: "system", content: msg[:content] }
+            { role: "system", content: content }
          end
        end

@ -63,11 +77,19 @@ module DiscourseAi
        end

        def tool_call_msg(msg)
-          tools_dialect.from_raw_tool_call(msg)
+          if disable_native_tools?
+            super
+          else
+            tools_dialect.from_raw_tool_call(msg)
+          end
        end

        def tool_msg(msg)
-          tools_dialect.from_raw_tool(msg)
+          if disable_native_tools?
+            super
+          else
+            tools_dialect.from_raw_tool(msg)
+          end
        end

        def user_msg(msg)
--- a/lib/completions/dialects/dialect.rb
+++ b/lib/completions/dialects/dialect.rb
@ -168,7 +168,7 @@ module DiscourseAi
          raise NotImplemented
        end

-        def assistant_msg(msg)
+        def model_msg(msg)
          raise NotImplemented
        end

@ -177,11 +177,15 @@ module DiscourseAi
        end

        def tool_call_msg(msg)
-          { role: "assistant", content: tools_dialect.from_raw_tool_call(msg) }
+          new_content = tools_dialect.from_raw_tool_call(msg)
+          msg = msg.merge(content: new_content)
+          model_msg(msg)
        end

        def tool_msg(msg)
-          { role: "user", content: tools_dialect.from_raw_tool(msg) }
+          new_content = tools_dialect.from_raw_tool(msg)
+          msg = msg.merge(content: new_content)
+          user_msg(msg)
        end
      end
    end
--- a/lib/completions/dialects/gemini.rb
+++ b/lib/completions/dialects/gemini.rb
@ -11,7 +11,7 @@ module DiscourseAi
        end

        def native_tool_support?
-          true
+          !llm_model.lookup_custom_param("disable_native_tools")
        end

        def translate
@ -84,10 +84,16 @@ module DiscourseAi
        end

        def system_msg(msg)
+          content = msg[:content]
+
+          if !native_tool_support? && tools_dialect.instructions.present?
+            content = content.to_s + "\n\n#{tools_dialect.instructions}"
+          end
+
          if beta_api?
-            { role: "system", content: msg[:content] }
+            { role: "system", content: content }
          else
-            { role: "user", parts: { text: msg[:content] } }
+            { role: "user", parts: { text: content } }
          end
        end

@ -125,35 +131,43 @@ module DiscourseAi
        end

        def tool_call_msg(msg)
-          call_details = JSON.parse(msg[:content], symbolize_names: true)
-          part = {
-            functionCall: {
-              name: msg[:name] || call_details[:name],
-              args: call_details[:arguments],
-            },
-          }
+          if native_tool_support?
+            call_details = JSON.parse(msg[:content], symbolize_names: true)
+            part = {
+              functionCall: {
+                name: msg[:name] || call_details[:name],
+                args: call_details[:arguments],
+              },
+            }

-          if beta_api?
-            { role: "model", parts: [part] }
+            if beta_api?
+              { role: "model", parts: [part] }
+            else
+              { role: "model", parts: part }
+            end
          else
-            { role: "model", parts: part }
+            super
          end
        end

        def tool_msg(msg)
-          part = {
-            functionResponse: {
-              name: msg[:name] || msg[:id],
-              response: {
-                content: msg[:content],
+          if native_tool_support?
+            part = {
+              functionResponse: {
+                name: msg[:name] || msg[:id],
+                response: {
+                  content: msg[:content],
+                },
              },
-            },
-          }
+            }

-          if beta_api?
-            { role: "function", parts: [part] }
+            if beta_api?
+              { role: "function", parts: [part] }
+            else
+              { role: "function", parts: part }
+            end
          else
-            { role: "function", parts: part }
+            super
          end
        end
      end
--- a/lib/completions/dialects/ollama.rb
+++ b/lib/completions/dialects/ollama.rb
@ -37,11 +37,19 @@ module DiscourseAi
        end

        def tool_call_msg(msg)
-          tools_dialect.from_raw_tool_call(msg)
+          if enable_native_tool?
+            tools_dialect.from_raw_tool_call(msg)
+          else
+            super
+          end
        end

        def tool_msg(msg)
-          tools_dialect.from_raw_tool(msg)
+          if enable_native_tool?
+            tools_dialect.from_raw_tool(msg)
+          else
+            super
+          end
        end

        def system_msg(msg)
--- a/lib/completions/dialects/xml_tools.rb
+++ b/lib/completions/dialects/xml_tools.rb
@ -118,8 +118,9 @@ module DiscourseAi
            If you wish to call multiple function in one reply, wrap multiple <invoke>
            block in a single <function_calls> block.

-            Always prefer to lead with tool calls, if you need to execute any.
-            Avoid all niceties prior to tool calls, Eg: "Let me look this up for you.." etc.
+            - Always prefer to lead with tool calls, if you need to execute any.
+            - Avoid all niceties prior to tool calls, Eg: "Let me look this up for you.." etc.
+            - DO NOT encode HTML entities in tool calls. You may use <![CDATA[...]]> for encoding if required.
            Here are the complete list of tools available:
          TEXT
        end
--- a/lib/completions/endpoints/anthropic.rb
+++ b/lib/completions/endpoints/anthropic.rb
@ -32,7 +32,11 @@ module DiscourseAi
              llm_model.name
            end

-          options = { model: mapped_model, max_tokens: 3_000 }
+          # Note: Anthropic requires this param
+          max_tokens = 4096
+          max_tokens = 8192 if mapped_model.match?(/3.5/)
+
+          options = { model: mapped_model, max_tokens: max_tokens }

          options[:stop_sequences] = ["</function_calls>"] if !dialect.native_tool_support? &&
            dialect.prompt.has_tools?
--- a/lib/completions/endpoints/base.rb
+++ b/lib/completions/endpoints/base.rb
@ -96,8 +96,10 @@ module DiscourseAi
                raise CompletionFailed, response.body
              end

-              xml_tool_processor = XmlToolProcessor.new if xml_tools_enabled? &&
-                dialect.prompt.has_tools?
+              xml_tool_processor =
+                XmlToolProcessor.new(
+                  partial_tool_calls: partial_tool_calls,
+                ) if xml_tools_enabled? && dialect.prompt.has_tools?

              to_strip = xml_tags_to_strip(dialect)
              xml_stripper =
--- a/lib/completions/endpoints/gemini.rb
+++ b/lib/completions/endpoints/gemini.rb
@ -58,7 +58,9 @@ module DiscourseAi
        end

        def prepare_payload(prompt, model_params, dialect)
-          tools = dialect.tools
+          @native_tool_support = dialect.native_tool_support?
+
+          tools = dialect.tools if @native_tool_support

          payload = default_options.merge(contents: prompt[:messages])
          payload[:systemInstruction] = {
@ -144,6 +146,7 @@ module DiscourseAi

        def decode(chunk)
          json = JSON.parse(chunk, symbolize_names: true)
+
          idx = -1
          json
            .dig(:candidates, 0, :content, :parts)
@ -168,30 +171,28 @@ module DiscourseAi

        def decode_chunk(chunk)
          @tool_index ||= -1
-
          streaming_decoder
            .decode(chunk)
            .map do |parsed|
              update_usage(parsed)
-              parsed
-                .dig(:candidates, 0, :content, :parts)
-                .map do |part|
-                  if part[:text]
-                    part = part[:text]
-                    if part != ""
-                      part
-                    else
-                      nil
-                    end
-                  elsif part[:functionCall]
-                    @tool_index += 1
-                    ToolCall.new(
-                      id: "tool_#{@tool_index}",
-                      name: part[:functionCall][:name],
-                      parameters: part[:functionCall][:args],
-                    )
+              parts = parsed.dig(:candidates, 0, :content, :parts)
+              parts&.map do |part|
+                if part[:text]
+                  part = part[:text]
+                  if part != ""
+                    part
+                  else
+                    nil
                  end
+                elsif part[:functionCall]
+                  @tool_index += 1
+                  ToolCall.new(
+                    id: "tool_#{@tool_index}",
+                    name: part[:functionCall][:name],
+                    parameters: part[:functionCall][:args],
+                  )
                end
+              end
            end
            .flatten
            .compact
@ -223,7 +224,7 @@ module DiscourseAi
        end

        def xml_tools_enabled?
-          false
+          !@native_tool_support
        end
      end
    end
--- a/lib/completions/endpoints/open_ai.rb
+++ b/lib/completions/endpoints/open_ai.rb
@ -36,14 +36,8 @@ module DiscourseAi
          partial_tool_calls: false,
          &blk
        )
-          if dialect.respond_to?(:is_gpt_o?) && dialect.is_gpt_o? && block_given?
-            # we need to disable streaming and simulate it
-            blk.call "", lambda { |*| }
-            response = super(dialect, user, model_params, feature_name: feature_name, &nil)
-            blk.call response, lambda { |*| }
-          else
-            super
-          end
+          @disable_native_tools = dialect.disable_native_tools?
+          super
        end

        private
@ -69,10 +63,17 @@ module DiscourseAi
            # We'll fallback to guess this using the tokenizer.
            payload[:stream_options] = { include_usage: true } if llm_model.provider == "open_ai"
          end
-          if dialect.tools.present?
-            payload[:tools] = dialect.tools
-            if dialect.tool_choice.present?
-              payload[:tool_choice] = { type: "function", function: { name: dialect.tool_choice } }
+          if !xml_tools_enabled?
+            if dialect.tools.present?
+              payload[:tools] = dialect.tools
+              if dialect.tool_choice.present?
+                payload[:tool_choice] = {
+                  type: "function",
+                  function: {
+                    name: dialect.tool_choice,
+                  },
+                }
+              end
            end
          end
          payload
@ -121,7 +122,7 @@ module DiscourseAi
        end

        def xml_tools_enabled?
-          false
+          !!@disable_native_tools
        end

        private
--- a/lib/completions/llm.rb
+++ b/lib/completions/llm.rb
@ -31,9 +31,8 @@ module DiscourseAi
                      tokens: 200_000,
                      display_name: "Claude 3.5 Sonnet",
                    },
+                    { name: "claude-3-5-haiku", tokens: 200_000, display_name: "Claude 3.5 Haiku" },
                    { name: "claude-3-opus", tokens: 200_000, display_name: "Claude 3 Opus" },
-                    { name: "claude-3-sonnet", tokens: 200_000, display_name: "Claude 3 Sonnet" },
-                    { name: "claude-3-haiku", tokens: 200_000, display_name: "Claude 3 Haiku" },
                  ],
                  tokenizer: DiscourseAi::Tokenizer::AnthropicTokenizer,
                  endpoint: "https://api.anthropic.com/v1/messages",
@ -63,6 +62,8 @@ module DiscourseAi
                {
                  id: "open_ai",
                  models: [
+                    { name: "o1-preview", tokens: 131_072, display_name: "o1" },
+                    { name: "o1-mini", tokens: 131_072, display_name: "o1 mini" },
                    { name: "gpt-4o", tokens: 131_072, display_name: "GPT-4 Omni" },
                    { name: "gpt-4o-mini", tokens: 131_072, display_name: "GPT-4 Omni Mini" },
                    { name: "gpt-4-turbo", tokens: 131_072, display_name: "GPT-4 Turbo" },
@ -71,6 +72,24 @@ module DiscourseAi
                  endpoint: "https://api.openai.com/v1/chat/completions",
                  provider: "open_ai",
                },
+                {
+                  id: "samba_nova",
+                  models: [
+                    {
+                      name: "Meta-Llama-3.1-8B-Instruct",
+                      tokens: 16_384,
+                      display_name: "Llama 3.1 8B",
+                    },
+                    {
+                      name: "Meta-Llama-3.1-70B-Instruct",
+                      tokens: 65_536,
+                      display_name: "Llama 3.1 70B",
+                    },
+                  ],
+                  tokenizer: DiscourseAi::Tokenizer::Llama3Tokenizer,
+                  endpoint: "https://api.sambanova.ai/v1/chat/completions",
+                  provider: "samba_nova",
+                },
              ]
            end
        end
--- a/lib/completions/tool_call.rb
+++ b/lib/completions/tool_call.rb
@ -6,6 +6,10 @@ module DiscourseAi
      attr_reader :id, :name, :parameters
      attr_accessor :partial

+      def partial?
+        !!@partial
+      end
+
      def initialize(id:, name:, parameters: nil)
        @id = id
        @name = name
--- a/lib/completions/tool_call_progress_tracker.rb
+++ b/lib/completions/tool_call_progress_tracker.rb
@ -16,7 +16,12 @@ module DiscourseAi
          @current_value = nil
        end

-        @parser.value { |v| tool_call.notify_progress(@current_key, v) if @current_key }
+        @parser.value do |v|
+          if @current_key
+            tool_call.notify_progress(@current_key, v)
+            @current_key = nil
+          end
+        end
      end

      def <<(json)
--- a/lib/completions/xml_tool_processor.rb
+++ b/lib/completions/xml_tool_processor.rb
@ -7,11 +7,13 @@
 module DiscourseAi
  module Completions
    class XmlToolProcessor
-      def initialize
+      def initialize(partial_tool_calls: false)
        @buffer = +""
        @function_buffer = +""
        @should_cancel = false
        @in_tool = false
+        @partial_tool_calls = partial_tool_calls
+        @partial_tools = [] if @partial_tool_calls
      end

      def <<(text)
@ -28,10 +30,10 @@ module DiscourseAi
          if @in_tool
            @function_buffer = @buffer[index..-1]
            text_index = text.rindex("<function_calls>")
-            result << text[0..text_index - 1].strip if text_index && text_index > 0
+            result << text[0..text_index - 1].rstrip if text_index && text_index > 0
          end
        else
-          @function_buffer << text
+          add_to_function_buffer(text)
        end

        if !@in_tool
@ -41,7 +43,7 @@ module DiscourseAi
              @function_buffer = text[split_index + 1..-1] || ""
              text = text[0..split_index] || ""
            else
-              @function_buffer << text
+              add_to_function_buffer(text)
              text = ""
            end
          else
@ -56,37 +58,25 @@ module DiscourseAi
          @should_cancel = true if text.include?("</function_calls>")
        end

+        if @should_notify_partial_tool
+          @should_notify_partial_tool = false
+          result << @partial_tools.last
+        end
+
        result
      end

      def finish
        return [] if @function_buffer.blank?

-        xml = Nokogiri::HTML5.fragment(@function_buffer)
-        normalize_function_ids!(xml)
-        last_invoke = xml.at("invoke:last")
-        if last_invoke
-          last_invoke.next_sibling.remove while last_invoke.next_sibling
-          xml.at("invoke:last").add_next_sibling("\n") if !last_invoke.next_sibling
+        idx = -1
+        parse_malformed_xml(@function_buffer).map do |tool|
+          ToolCall.new(
+            id: "tool_#{idx += 1}",
+            name: tool[:tool_name],
+            parameters: tool[:parameters],
+          )
        end
-
-        xml
-          .css("invoke")
-          .map do |invoke|
-            tool_name = invoke.at("tool_name").content.force_encoding("UTF-8")
-            tool_id = invoke.at("tool_id").content.force_encoding("UTF-8")
-            parameters = {}
-            invoke
-              .at("parameters")
-              &.children
-              &.each do |node|
-                next if node.text?
-                name = node.name
-                value = node.content.to_s
-                parameters[name.to_sym] = value.to_s.force_encoding("UTF-8")
-              end
-            ToolCall.new(id: tool_id, name: tool_name, parameters: parameters)
-          end
      end

      def should_cancel?
@ -95,6 +85,105 @@ module DiscourseAi

      private

+      def add_to_function_buffer(text)
+        @function_buffer << text
+        detect_partial_tool_calls(@function_buffer, text) if @partial_tool_calls
+      end
+
+      def detect_partial_tool_calls(buffer, delta)
+        parse_partial_tool_call(buffer)
+      end
+
+      def parse_partial_tool_call(buffer)
+        match =
+          buffer
+            .scan(
+              %r{
+      <invoke>
+        \s*
+        <tool_name>
+          ([^<]+)
+        </tool_name>
+        \s*
+        <parameters>
+          (.*?)
+        (</parameters>|\Z)
+        }mx,
+            )
+            .to_a
+            .last
+
+        if match
+          params = partial_parse_params(match[1])
+          if params.present?
+            current_tool = @partial_tools.last
+            if !current_tool || current_tool.name != match[0].strip
+              current_tool =
+                ToolCall.new(
+                  id: "tool_#{@partial_tools.length}",
+                  name: match[0].strip,
+                  parameters: params,
+                )
+              @partial_tools << current_tool
+              current_tool.partial = true
+              @should_notify_partial_tool = true
+            end
+
+            if current_tool.parameters != params
+              current_tool.parameters = params
+              @should_notify_partial_tool = true
+            end
+          end
+        end
+      end
+
+      def partial_parse_params(params)
+        params
+          .scan(%r{
+      <([^>]+)>
+        (.*?)
+      (</\1>|\Z)
+    }mx)
+          .each_with_object({}) do |(name, value), hash|
+            next if "<![CDATA[".start_with?(value)
+            hash[name.to_sym] = value.gsub(/^<!\[CDATA\[|\]\]>$/, "")
+          end
+      end
+
+      def parse_malformed_xml(input)
+        input
+          .scan(
+            %r{
+      <invoke>
+        \s*
+        <tool_name>
+          ([^<]+)
+        </tool_name>
+        \s*
+        <parameters>
+          (.*?)
+        </parameters>
+        \s*
+      </invoke>
+    }mx,
+          )
+          .map do |tool_name, params|
+            {
+              tool_name: tool_name.strip,
+              parameters:
+                params
+                  .scan(%r{
+          <([^>]+)>
+            (.*?)
+          </\1>
+        }mx)
+                  .each_with_object({}) do |(name, value), hash|
+                    hash[name.to_sym] = value.gsub(/^<!\[CDATA\[|\]\]>$/, "")
+                  end,
+            }
+          end
+      end
+
      def normalize_function_ids!(function_buffer)
        function_buffer
          .css("invoke")
--- a/plugin.rb
+++ b/plugin.rb
@ -39,6 +39,8 @@ register_asset "stylesheets/modules/llms/common/ai-llms-editor.scss"

 register_asset "stylesheets/modules/ai-bot/common/ai-tools.scss"

+register_asset "stylesheets/modules/ai-bot/common/ai-artifact.scss"
+
 module ::DiscourseAi
  PLUGIN_NAME = "discourse-ai"
 end
@ -48,6 +50,10 @@ Rails.autoloaders.main.push_dir(File.join(__dir__, "lib"), namespace: ::Discours
 require_relative "lib/engine"

 after_initialize do
+  if defined?(Rack::MiniProfiler)
+    Rack::MiniProfiler.config.skip_paths << "/discourse-ai/ai-bot/artifacts"
+  end
+
  # do not autoload this cause we may have no namespace
  require_relative "discourse_automation/llm_triage"
  require_relative "discourse_automation/llm_report"
--- a/spec/fixtures/bot/openai_artifact_call.txt
+++ b/spec/fixtures/bot/openai_artifact_call.txt
@ -0,0 +1,299 @@
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"role":"assistant","content":null,"tool_calls":[{"index":0,"id":"call_yH3ixdEz4wvSuK8ei3gNYwk3","type":"function","function":{"name":"create_artifact","arguments":""}}],"refusal":null},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"{\""}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"name"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\":\""}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"Five"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"Lines"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"HTML"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"CSS"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"JS"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"Hello"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"World"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\",\""}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"html"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"_body"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\":\""}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"<h"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"1"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":">Hello"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"</"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"h"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"1"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":">\\"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"n"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"<p"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":">"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"World"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"!</"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"p"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":">\\"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"n"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"<button"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" onclick"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"=\\\""}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"show"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"Message"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"()"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\\\">"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"G"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"reet"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"</"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"button"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":">\\"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"n"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"<div"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" id"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"=\\\""}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"message"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\\\"></"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"div"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":">\\"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"n"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"<hr"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":">"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\",\""}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"css"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\":\""}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"body"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" {"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" margin"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":":"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" "}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"0"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":";"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" }\\"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"nh"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"1"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" {"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" color"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":":"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" blue"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":";"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" }\\"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"np"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" {"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" font"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"-size"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":":"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" "}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"20"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"px"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":";"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" }\\"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"n"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"button"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" {"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" padding"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":":"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" "}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"10"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"px"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":";"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" }\\"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"n"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"hr"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" {"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" border"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":":"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" "}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"1"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"px"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" solid"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" #"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"ccc"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":";"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" }"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\",\""}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"js"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\":\""}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"function"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" show"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"Message"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"()"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" {\\"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"n"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" "}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" var"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" message"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"Div"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" ="}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" document"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":".get"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"Element"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"By"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"Id"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"('"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"message"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"');"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\\"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"n"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" "}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" message"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"Div"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":".text"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"Content"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" ="}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" '"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"Hello"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":","}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" World"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"!"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"';"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\\"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"n"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"}\\"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"n"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\"}"}}]},"logprobs":null,"finish_reason":null}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[{"index":0,"delta":{},"logprobs":null,"finish_reason":"tool_calls"}],"usage":null}
+
+data: {"id":"chatcmpl-ATimVYagKnCWQ0VXY0Hn2SDjRuN6B","object":"chat.completion.chunk","created":1731647015,"model":"gpt-4o-2024-08-06","system_fingerprint":"fp_45cf54deae","choices":[],"usage":{"prompt_tokens":735,"completion_tokens":156,"total_tokens":891,"prompt_tokens_details":{"cached_tokens":0,"audio_tokens":0},"completion_tokens_details":{"reasoning_tokens":0,"audio_tokens":0,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}}}
+
+data: [DONE]
+
+
--- a/spec/lib/completions/endpoints/anthropic_spec.rb
+++ b/spec/lib/completions/endpoints/anthropic_spec.rb
@ -186,7 +186,7 @@ RSpec.describe DiscourseAi::Completions::Endpoints::Anthropic do

    expected_body = {
      model: "claude-3-opus-20240229",
-      max_tokens: 3000,
+      max_tokens: 4096,
      messages: [{ role: "user", content: "user1: hello" }],
      system: "You are hello bot",
      stream: true,
@ -278,7 +278,7 @@ RSpec.describe DiscourseAi::Completions::Endpoints::Anthropic do

    request_body = {
      model: "claude-3-opus-20240229",
-      max_tokens: 3000,
+      max_tokens: 4096,
      messages: [
        {
          role: "user",
@ -376,7 +376,7 @@ RSpec.describe DiscourseAi::Completions::Endpoints::Anthropic do

    expected_body = {
      model: "claude-3-opus-20240229",
-      max_tokens: 3000,
+      max_tokens: 4096,
      messages: [{ role: "user", content: "user1: hello" }],
      system: "You are hello bot",
    }
--- a/spec/lib/completions/endpoints/gemini_spec.rb
+++ b/spec/lib/completions/endpoints/gemini_spec.rb
@ -324,6 +324,37 @@ RSpec.describe DiscourseAi::Completions::Endpoints::Gemini do
    expect(log.response_tokens).to eq(4)
  end

+  it "Can correctly handle malformed responses" do
+    response = <<~TEXT
+      data: {"candidates": [{"content": {"parts": [{"text": "Certainly"}],"role": "model"}}],"usageMetadata": {"promptTokenCount": 399,"totalTokenCount": 399},"modelVersion": "gemini-1.5-pro-002"}
+
+      data: {"candidates": [{"content": {"parts": [{"text": "! I'll create a simple \\"Hello, World!\\" page where each letter"}],"role": "model"},"safetyRatings": [{"category": "HARM_CATEGORY_HATE_SPEECH","probability": "NEGLIGIBLE"},{"category": "HARM_CATEGORY_DANGEROUS_CONTENT","probability": "NEGLIGIBLE"},{"category": "HARM_CATEGORY_HARASSMENT","probability": "NEGLIGIBLE"},{"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT","probability": "NEGLIGIBLE"}]}],"usageMetadata": {"promptTokenCount": 399,"totalTokenCount": 399},"modelVersion": "gemini-1.5-pro-002"}
+
+      data: {"candidates": [{"content": {"parts": [{"text": " has a different color using inline styles for simplicity.  Each letter will be wrapped"}],"role": "model"},"safetyRatings": [{"category": "HARM_CATEGORY_HATE_SPEECH","probability": "NEGLIGIBLE"},{"category": "HARM_CATEGORY_DANGEROUS_CONTENT","probability": "NEGLIGIBLE"},{"category": "HARM_CATEGORY_HARASSMENT","probability": "NEGLIGIBLE"},{"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT","probability": "NEGLIGIBLE"}]}],"usageMetadata": {"promptTokenCount": 399,"totalTokenCount": 399},"modelVersion": "gemini-1.5-pro-002"}
+
+      data: {"candidates": [{"content": {"parts": [{"text": ""}],"role": "model"},"finishReason": "STOP"}],"usageMetadata": {"promptTokenCount": 399,"candidatesTokenCount": 191,"totalTokenCount": 590},"modelVersion": "gemini-1.5-pro-002"}
+
+      data: {"candidates": [{"finishReason": "MALFORMED_FUNCTION_CALL"}],"usageMetadata": {"promptTokenCount": 399,"candidatesTokenCount": 191,"totalTokenCount": 590},"modelVersion": "gemini-1.5-pro-002"}
+
+    TEXT
+
+    llm = DiscourseAi::Completions::Llm.proxy("custom:#{model.id}")
+    url = "#{model.url}:streamGenerateContent?alt=sse&key=123"
+
+    output = []
+
+    stub_request(:post, url).to_return(status: 200, body: response)
+    llm.generate("Hello", user: user) { |partial| output << partial }
+
+    expect(output).to eq(
+      [
+        "Certainly",
+        "! I'll create a simple \"Hello, World!\" page where each letter",
+        " has a different color using inline styles for simplicity.  Each letter will be wrapped",
+      ],
+    )
+  end
+
  it "Can correctly handle streamed responses even if they are chunked badly" do
    data = +""
    data << "da|ta: |"
--- a/spec/lib/completions/endpoints/open_ai_spec.rb
+++ b/spec/lib/completions/endpoints/open_ai_spec.rb
@ -53,11 +53,10 @@ class OpenAiMock < EndpointMock
    }.to_json
  end

-  def stub_raw(chunks)
-    WebMock.stub_request(:post, "https://api.openai.com/v1/chat/completions").to_return(
-      status: 200,
-      body: chunks,
-    )
+  def stub_raw(chunks, body_blk: nil)
+    stub = WebMock.stub_request(:post, "https://api.openai.com/v1/chat/completions")
+    stub.with(body: body_blk) if body_blk
+    stub.to_return(status: 200, body: chunks)
  end

  def stub_streamed_response(prompt, deltas, tool_call: false)
@ -391,6 +390,59 @@ RSpec.describe DiscourseAi::Completions::Endpoints::OpenAi do
  end

  describe "#perform_completion!" do
+    context "when using XML tool calls format" do
+      let(:xml_tool_call_response) { <<~XML }
+        <function_calls>
+        <invoke>
+        <tool_name>get_weather</tool_name>
+        <parameters>
+        <location>Sydney</location>
+        <unit>c</unit>
+        </parameters>
+        </invoke>
+        </function_calls>
+      XML
+
+      it "parses XML tool calls" do
+        response = {
+          id: "chatcmpl-6sZfAb30Rnv9Q7ufzFwvQsMpjZh8S",
+          object: "chat.completion",
+          created: 1_678_464_820,
+          model: "gpt-3.5-turbo-0301",
+          usage: {
+            prompt_tokens: 8,
+            completion_tokens: 13,
+            total_tokens: 499,
+          },
+          choices: [
+            {
+              message: {
+                role: "assistant",
+                content: xml_tool_call_response,
+              },
+              finish_reason: "stop",
+              index: 0,
+            },
+          ],
+        }.to_json
+
+        endpoint.llm_model.update!(provider_params: { disable_native_tools: true })
+        body = nil
+        open_ai_mock.stub_raw(response, body_blk: proc { |inner_body| body = inner_body })
+
+        dialect = compliance.dialect(prompt: compliance.generic_prompt(tools: tools))
+        tool_call = endpoint.perform_completion!(dialect, user)
+
+        body_parsed = JSON.parse(body, symbolize_names: true)
+        expect(body_parsed[:tools]).to eq(nil)
+
+        expect(body_parsed[:messages][0][:content]).to include("<function_calls>")
+
+        expect(tool_call.name).to eq("get_weather")
+        expect(tool_call.parameters).to eq({ location: "Sydney", unit: "c" })
+      end
+    end
+
    context "when using regular mode" do
      context "with simple prompts" do
        it "completes a trivial prompt and logs the response" do
@ -571,6 +623,42 @@ TEXT
          end
        end

+        it "properly handles multiple params in partial tool calls" do
+          # this is not working and it is driving me nuts so I will use a sledghammer
+          # text = plugin_file_from_fixtures("openai_artifact_call.txt", "bot")
+
+          path = File.join(__dir__, "../../../fixtures/bot", "openai_artifact_call.txt")
+          text = File.read(path)
+
+          partials = []
+          open_ai_mock.with_chunk_array_support do
+            open_ai_mock.stub_raw(text.scan(/.*\n/))
+
+            dialect = compliance.dialect(prompt: compliance.generic_prompt(tools: tools))
+            endpoint.perform_completion!(dialect, user, partial_tool_calls: true) do |partial|
+              partials << partial.dup
+            end
+          end
+
+          expect(partials.compact.length).to eq(128)
+
+          params =
+            partials
+              .map { |p| p.parameters if p.is_a?(DiscourseAi::Completions::ToolCall) && p.partial? }
+              .compact
+
+          lengths = {}
+          params.each do |p|
+            p.each do |k, v|
+              if lengths[k] && lengths[k] > v.length
+                expect(lengths[k]).to be > v.length
+              else
+                lengths[k] = v.length
+              end
+            end
+          end
+        end
+
        it "properly handles spaces in tools payload and partial tool calls" do
          raw_data = <<~TEXT.strip
            data: {"choices":[{"index":0,"delta":{"role":"assistant","content":null,"tool_calls":[{"index":0,"id":"func_id","type":"function","function":{"name":"go|ogle","arg|uments":""}}]}}]}
--- a/spec/lib/completions/xml_tool_processor_spec.rb
+++ b/spec/lib/completions/xml_tool_processor_spec.rb
@ -12,12 +12,84 @@ RSpec.describe DiscourseAi::Completions::XmlToolProcessor do
    expect(processor.should_cancel?).to eq(false)
  end

-  it "is usable for simple single message mode" do
+  it "can handle partial tool calls" do
+    processor = DiscourseAi::Completions::XmlToolProcessor.new(partial_tool_calls: true)
+
+    xml = (<<~XML).strip
+      <function|_calls>
+      <invoke>
+      <tool_name>h|ell|o<|/tool_name>
+      <parameters>
+       <hello>wo|r|ld</hello>
+      </parameters>
+      </invoke>
+      <invoke>
+      <tool_name>tool|2</tool_name>
+      <parameters>
+        <param>v|alue</param>
+        <para|m2><![CDA|TA[va|lue2]]></param2>
+      </parame|ters>
+      </invoke>
+    XML
+
+    result = []
+
+    xml.split("|").each { |part| result << (processor << part).map(&:dup) }
+
+    result << (processor.finish)
+    result.flatten!
+
+    tool1_params =
+      result
+        .select do |r|
+          r.is_a?(DiscourseAi::Completions::ToolCall) && r.name == "hello" && r.partial
+        end
+        .map(&:parameters)
+
+    expect(tool1_params).to eq([{ hello: "wo" }, { hello: "wor" }, { hello: "world" }])
+
+    tool2_params =
+      result
+        .select do |r|
+          r.is_a?(DiscourseAi::Completions::ToolCall) && r.name == "tool2" && r.partial
+        end
+        .map(&:parameters)
+
+    expect(tool2_params).to eq(
+      [
+        { param: "v" },
+        { param: "value" },
+        { param: "value", param2: "va" },
+        { param: "value", param2: "value2" },
+      ],
+    )
+  end
+
+  it "can handle mix and match xml cause tool llms may not encode" do
    xml = (<<~XML).strip
-      hello
      <function_calls>
      <invoke>
      <tool_name>hello</tool_name>
+      <parameters>
+       <hello>world <sam>sam</sam></hello>
+       <test><![CDATA[</h1>\n</div>\n]]></test>
+      </parameters>
+      </invoke>
+    XML
+
+    result = []
+    result << (processor << xml)
+    result << (processor.finish)
+
+    tool_call = result.last.first
+    expect(tool_call.parameters).to eq(hello: "world <sam>sam</sam>", test: "</h1>\n</div>\n")
+  end
+
+  it "is usable for simple single message mode" do
+    xml = (<<~XML)
+       world <function_calls>
+      <invoke>
+      <tool_name>hello</tool_name>
      <parameters>
       <hello>world</hello>
       <test>value</test>
@ -26,6 +98,7 @@ RSpec.describe DiscourseAi::Completions::XmlToolProcessor do
    XML

    result = []
+    result << (processor << "hello")
    result << (processor << xml)
    result << (processor.finish)

@ -38,7 +111,7 @@ RSpec.describe DiscourseAi::Completions::XmlToolProcessor do
          test: "value",
        },
      )
-    expect(result).to eq([["hello"], [tool_call]])
+    expect(result).to eq([["hello"], [" world"], [tool_call]])
    expect(processor.should_cancel?).to eq(false)
  end

@ -149,8 +222,7 @@ RSpec.describe DiscourseAi::Completions::XmlToolProcessor do
    result << (processor.finish)

    # Should just do its best to parse the XML
-    tool_call =
-      DiscourseAi::Completions::ToolCall.new(id: "tool_0", name: "test", parameters: { param: "" })
+    tool_call = DiscourseAi::Completions::ToolCall.new(id: "tool_0", name: "test", parameters: {})
    expect(result).to eq([["text"], [tool_call]])
  end

--- a/spec/lib/modules/ai_bot/personas/persona_spec.rb
+++ b/spec/lib/modules/ai_bot/personas/persona_spec.rb
@ -46,6 +46,7 @@ RSpec.describe DiscourseAi::AiBot::Personas::Persona do
    }
  end

+  fab!(:admin)
  fab!(:user)
  fab!(:upload)

@ -96,29 +97,6 @@ RSpec.describe DiscourseAi::AiBot::Personas::Persona do
  end

  it "enforces enums" do
-    xml = <<~XML
-      <function_calls>
-        <invoke>
-        <tool_name>search</tool_name>
-        <tool_id>call_JtYQMful5QKqw97XFsHzPweB</tool_id>
-        <parameters>
-        <max_posts>"3.2"</max_posts>
-        <status>cow</status>
-        <foo>bar</foo>
-        </parameters>
-        </invoke>
-        <invoke>
-        <tool_name>search</tool_name>
-        <tool_id>call_JtYQMful5QKqw97XFsHzPweB</tool_id>
-        <parameters>
-        <max_posts>"3.2"</max_posts>
-        <status>open</status>
-        <foo>bar</foo>
-        </parameters>
-        </invoke>
-      </function_calls>
-    XML
-
    tool_call =
      DiscourseAi::Completions::ToolCall.new(
        name: "search",
@ -273,11 +251,27 @@ RSpec.describe DiscourseAi::AiBot::Personas::Persona do
        ],
      )

+      # it should allow staff access to WebArtifactCreator
+      expect(DiscourseAi::AiBot::Personas::Persona.all(user: admin)).to eq(
+        [
+          DiscourseAi::AiBot::Personas::General,
+          DiscourseAi::AiBot::Personas::Artist,
+          DiscourseAi::AiBot::Personas::Creative,
+          DiscourseAi::AiBot::Personas::DiscourseHelper,
+          DiscourseAi::AiBot::Personas::GithubHelper,
+          DiscourseAi::AiBot::Personas::Researcher,
+          DiscourseAi::AiBot::Personas::SettingsExplorer,
+          DiscourseAi::AiBot::Personas::SqlHelper,
+          DiscourseAi::AiBot::Personas::WebArtifactCreator,
+        ],
+      )
+
      # omits personas if key is missing
      SiteSetting.ai_stability_api_key = ""
      SiteSetting.ai_google_custom_search_api_key = ""
+      SiteSetting.ai_artifact_security = "disabled"

-      expect(DiscourseAi::AiBot::Personas::Persona.all(user: user)).to contain_exactly(
+      expect(DiscourseAi::AiBot::Personas::Persona.all(user: admin)).to contain_exactly(
        DiscourseAi::AiBot::Personas::General,
        DiscourseAi::AiBot::Personas::SqlHelper,
        DiscourseAi::AiBot::Personas::SettingsExplorer,
--- a/spec/lib/modules/ai_bot/playground_spec.rb
+++ b/spec/lib/modules/ai_bot/playground_spec.rb
@ -791,11 +791,12 @@ RSpec.describe DiscourseAi::AiBot::Playground do
        expect(done_signal.data[:cooked]).to eq(reply.cooked)

        expect(messages.first.data[:raw]).to eq("")
-        messages[1..-1].each_with_index do |m, idx|
-          expect(m.data[:raw]).to eq(expected_bot_response[0..idx])
-        end

        expect(reply.cooked).to eq(PrettyText.cook(expected_bot_response))
+
+        messages[1..-1].each do |m|
+          expect(expected_bot_response.start_with?(m.data[:raw])).to eq(true)
+        end
      end
    end

--- a/spec/lib/modules/ai_bot/tools/create_artifact_spec.rb
+++ b/spec/lib/modules/ai_bot/tools/create_artifact_spec.rb
@ -0,0 +1,34 @@
+#frozen_string_literal: true
+
+RSpec.describe DiscourseAi::AiBot::Tools::CreateArtifact do
+  fab!(:llm_model)
+  let(:bot_user) { DiscourseAi::AiBot::EntryPoint.find_user_from_model(llm_model.name) }
+  let(:llm) { DiscourseAi::Completions::Llm.proxy("custom:#{llm_model.id}") }
+
+  before { SiteSetting.ai_bot_enabled = true }
+
+  describe "#process" do
+    it "can correctly handle partial updates" do
+      tool = described_class.new({}, bot_user: bot_user, llm: llm)
+
+      tool.parameters = { css: "a { }" }
+      tool.partial_invoke
+
+      expect(tool.custom_raw).to eq("### CSS\n\n```css\na { }\n```")
+
+      tool.parameters = { css: "a { }", html_body: "hello" }
+      tool.partial_invoke
+
+      expect(tool.custom_raw).to eq(
+        "### CSS\n\n```css\na { }\n```\n\n### HTML\n\n```html\nhello\n```",
+      )
+
+      tool.parameters = { css: "a { }", html_body: "hello world" }
+      tool.partial_invoke
+
+      expect(tool.custom_raw).to eq(
+        "### CSS\n\n```css\na { }\n```\n\n### HTML\n\n```html\nhello world\n```",
+      )
+    end
+  end
+end
--- a/spec/requests/ai_bot/shared_ai_conversations_spec.rb
+++ b/spec/requests/ai_bot/shared_ai_conversations_spec.rb
@ -86,6 +86,77 @@ RSpec.describe DiscourseAi::AiBot::SharedAiConversationsController do
        expect(response).to have_http_status(:success)
      end

+      context "when ai artifacts are in lax mode" do
+        before { SiteSetting.ai_artifact_security = "lax" }
+
+        it "properly shares artifacts" do
+          first_post = user_pm_share.posts.first
+
+          artifact_not_allowed =
+            AiArtifact.create!(
+              user: bot_user,
+              post: Fabricate(:private_message_post),
+              name: "test",
+              html: "<div>test</div>",
+            )
+
+          artifact =
+            AiArtifact.create!(
+              user: bot_user,
+              post: first_post,
+              name: "test",
+              html: "<div>test</div>",
+            )
+
+          # lets log out and see we can not access the artifacts
+          delete "/session/#{user.id}"
+
+          get artifact.url
+          expect(response).to have_http_status(:not_found)
+
+          get artifact_not_allowed.url
+          expect(response).to have_http_status(:not_found)
+
+          sign_in(user)
+
+          first_post.update!(raw: <<~RAW)
+            This is a post with an artifact
+
+            <div class="ai-artifact" data-ai-artifact-id="#{artifact.id}"></div>
+            <div class="ai-artifact" data-ai-artifact-id="#{artifact_not_allowed.id}"></div>
+          RAW
+
+          post "#{path}.json", params: { topic_id: user_pm_share.id }
+          expect(response).to have_http_status(:success)
+
+          key = response.parsed_body["share_key"]
+
+          get "#{path}/#{key}"
+          expect(response).to have_http_status(:success)
+
+          expect(response.body).to include(artifact.url)
+          expect(response.body).to include(artifact_not_allowed.url)
+
+          # lets log out and see we can not access the artifacts
+          delete "/session/#{user.id}"
+
+          get artifact.url
+          expect(response).to have_http_status(:success)
+
+          get artifact_not_allowed.url
+          expect(response).to have_http_status(:not_found)
+
+          sign_in(user)
+          delete "#{path}/#{key}.json"
+          expect(response).to have_http_status(:success)
+
+          # we can not longer see it...
+          delete "/session/#{user.id}"
+          get artifact.url
+          expect(response).to have_http_status(:not_found)
+        end
+      end
+
      context "when secure uploads are enabled" do
        let(:upload_1) { Fabricate(:s3_image_upload, user: bot_user, secure: true) }
        let(:upload_2) { Fabricate(:s3_image_upload, user: bot_user, secure: true) }
--- a/spec/system/llms/ai_llm_spec.rb
+++ b/spec/system/llms/ai_llm_spec.rb
@ -11,7 +11,7 @@ RSpec.describe "Managing LLM configurations", type: :system, js: true do
  it "correctly sets defaults" do
    visit "/admin/plugins/discourse-ai/ai-llms"

-    find("[data-llm-id='anthropic-claude-3-haiku'] button").click()
+    find("[data-llm-id='anthropic-claude-3-5-haiku'] button").click()
    find("input.ai-llm-editor__api-key").fill_in(with: "abcd")
    find(".ai-llm-editor__enabled-chat-bot input").click
    find(".ai-llm-editor__save").click()
@ -23,9 +23,9 @@ RSpec.describe "Managing LLM configurations", type: :system, js: true do

    preset = DiscourseAi::Completions::Llm.presets.find { |p| p[:id] == "anthropic" }

-    model_preset = preset[:models].find { |m| m[:name] == "claude-3-haiku" }
+    model_preset = preset[:models].find { |m| m[:name] == "claude-3-5-haiku" }

-    expect(llm.name).to eq("claude-3-haiku")
+    expect(llm.name).to eq("claude-3-5-haiku")
    expect(llm.url).to eq(preset[:endpoint])
    expect(llm.tokenizer).to eq(preset[:tokenizer].to_s)
    expect(llm.max_prompt_tokens.to_i).to eq(model_preset[:tokens])