Merge pull request #107 from merefield/confirm_legal_non_post_urls

FEATURE: reject LLM responses with hallucinated non-post urls

lib/discourse_chatbot/bots/open_ai_bot_rag.rb

@@ -28,6 +28,7 @@ def get_response(prompt, opts)
       @inner_thoughts = []
       @posts_ids_found = []
       @topic_ids_found = []
+      @non_post_urls_found = []
 
       @chat_history += prompt
 
@@ -185,7 +186,7 @@ def generate_response(opts)
 
         if (['stop','length'].include?(finish_reason) && tools_calls.nil? || @inner_thoughts.length > 7)
           if iteration > 1 && SiteSetting.chatbot_url_integrity_check
-            if legal_urls?(res["choices"][0]["message"]["content"], @posts_ids_found, @topic_ids_found)
+            if legal_post_urls?(res["choices"][0]["message"]["content"], @posts_ids_found, @topic_ids_found) && legal_non_post_urls?(res["choices"][0]["message"]["content"], @non_post_urls_found)
               return res
             else
               @inner_thoughts << { role: 'user', content: I18n.t("chatbot.prompt.system.rag.illegal_urls") }
@@ -248,9 +249,10 @@ def handle_function_call(res, opts)
         tool_call_id = function_called["id"]
         if func_name == "local_forum_search"
           result_hash = call_function(func_name, args_str, opts)
-          result, post_ids_found, topic_ids_found = result_hash.values_at(:result, :post_ids_found, :topic_ids_found)
+          result, post_ids_found, topic_ids_found, non_post_urls_found = result_hash.values_at(:result, :post_ids_found, :topic_ids_found, :non_post_urls_found)
           @posts_ids_found = (@posts_ids_found.to_set | post_ids_found.to_set).to_a
           @topic_ids_found = (@topic_ids_found.to_set | topic_ids_found.to_set).to_a
+          @non_post_urls_found = (@non_post_urls_found.to_set | non_post_urls_found.to_set).to_a
         else
           result = call_function(func_name, args_str, opts)
         end
@@ -282,7 +284,7 @@ def call_function(func_name, args_str, opts)
       end
     end
 
-    def legal_urls?(res, post_ids_found, topic_ids_found)
+    def legal_post_urls?(res, post_ids_found, topic_ids_found)
       return true if res.blank?
 
       post_url_regex = ::DiscourseChatbot::POST_URL_REGEX
@@ -311,6 +313,22 @@ def legal_urls?(res, post_ids_found, topic_ids_found)
       true
     end
 
+    def legal_non_post_urls?(res, non_post_urls_found)
+      return true if res.blank?
+      non_post_url_regex = ::DiscourseChatbot::NON_POST_URL_REGEX
+
+      urls_in_text = res.scan(non_post_url_regex)
+
+      urls_in_text = urls_in_text.reject { |url| url.include?('/t/') }
+
+      urls_in_text.each do |url|
+        if !non_post_urls_found.include?(url)
+          return false
+        end
+      end
+      true
+    end
+
     private
 
     def image_url?(string)

lib/discourse_chatbot/functions/forum_search_function.rb

@@ -34,6 +34,7 @@ def process(args)
         top_topic_title_results = []
         post_ids_found = []
         topic_ids_found = []
+        non_post_urls_found = []
         query = args[parameters[0][:name]]
         number_of_posts = args[parameters[1][:name]].blank? ? 3 : args[parameters[1][:name]]
         number_of_posts = number_of_posts > SiteSetting.chatbot_forum_search_function_max_results ? SiteSetting.chatbot_forum_search_function_max_results : number_of_posts
@@ -103,18 +104,21 @@ def process(args)
               response += I18n.t("chatbot.prompt.function.forum_search.answer.topic.each.post", post_number: post_number, username: post.user.username, date: post.created_at, raw: post.raw)
 
               topic_ids_in_raw_urls_found, post_ids_in_raw_urls_found = find_post_and_topic_ids_from_raw_urls(post.raw)
+              non_post_urls_found = non_post_urls_found | find_other_urls(post.raw)
 
               topic_ids_found = topic_ids_found | topic_ids_in_raw_urls_found
               post_ids_found = post_ids_found | post_ids_in_raw_urls_found
 
               post_ids_found << post.id
+              topic_ids_found << post.topic_id
               post_number += 1
             end
           end
         else
           response = I18n.t("chatbot.prompt.function.forum_search.answer.post.summary", number_of_posts: number_of_posts)
           top_results.each_with_index do |result, index|
             current_post = ::Post.find(result[:post_id].to_i)
+
             score = result[:score]
             url = "https://#{Discourse.current_hostname}/t/slug/#{current_post.topic_id}/#{current_post.post_number}"
             raw = current_post.raw
@@ -127,10 +131,13 @@ def process(args)
             topic_ids_found = topic_ids_found | topic_ids_in_raw_urls_found
             post_ids_found = post_ids_found | post_ids_in_raw_urls_found
 
-            post_ids_found << current_post.id
+            non_post_urls_found = non_post_urls_found | find_other_urls(raw)
+
+            post_ids_found = post_ids_found | [current_post.id]
+            topic_ids_found = topic_ids_found | [current_post.topic_id]
           end
         end
-        { result: response, topic_ids_found: topic_ids_found, post_ids_found: post_ids_found }
+        { result: response, topic_ids_found: topic_ids_found, post_ids_found: post_ids_found, non_post_urls_found: non_post_urls_found}
       rescue StandardError => e
         Rails.logger.error("Chatbot: Error occurred while attempting to retrieve Forum Search results for query '#{query}': #{e.message}")
         { result: I18n.t("chatbot.prompt.function.forum_search.error", query: args[parameters[0][:name]]), topic_ids_found: [], post_ids_found: [] }
@@ -159,5 +166,11 @@ def find_post_and_topic_ids_from_raw_urls(raw)
 
       [topic_ids_found, post_ids_found]
     end
+
+    def find_other_urls(raw)
+      urls = raw.scan(::DiscourseChatbot::NON_POST_URL_REGEX)
+
+      urls.reject { |url| url.include?('/t/') }
+    end
   end
 end

plugin.rb

@@ -1,7 +1,7 @@
 # frozen_string_literal: true
 # name: discourse-chatbot
 # about: a plugin that allows you to have a conversation with a configurable chatbot in Discourse Chat, Topics and Private Messages
-# version: 0.9.37
+# version: 0.9.38
 # authors: merefield
 # url: https://github.com/merefield/discourse-chatbot
 
@@ -36,6 +36,7 @@ module ::DiscourseChatbot
 
   TOPIC_URL_REGEX = %r{\/t/[^/]+/(\d+)(?!\d|\/)}
   POST_URL_REGEX = %r{\/t/[^/]+/(\d+)/(\d+)(?!\d|\/)}
+  NON_POST_URL_REGEX = %r{\bhttps?:\/\/[^\s\/$.?#].[^\s)]*}
 
   def progress_debug_message(message)
     puts "Chatbot: #{message}" if SiteSetting.chatbot_enable_verbose_console_logging

spec/lib/bot/open_ai_agent_spec.rb

@@ -32,26 +32,34 @@
   end
 
   it "returns correct status for a response that includes and illegal topic id" do
-    result = rag.legal_urls?(res, post_ids_found, topic_ids_found)
+    result = rag.legal_post_urls?(res, post_ids_found, topic_ids_found)
 
     expect(result).to eq(false)
   end
 
   it "returns correct status for a response that includes a legal post id" do
     expect(post_1).to be_present
-    result = rag.legal_urls?(res_2, post_ids_found_2, topic_ids_found)
+    result = rag.legal_post_urls?(res_2, post_ids_found_2, topic_ids_found)
     expect(result).to eq(true)
   end
 
   it "correctly identifies a legal post id in a url in a response" do
-    expect(described_class.new({}).legal_urls?("hello /t/slug/112/2", [post_1.id], [topic_1.id])).to eq(true)
+    expect(described_class.new({}).legal_post_urls?("hello /t/slug/112/2", [post_1.id], [topic_1.id])).to eq(true)
   end
 
   it "correctly skips a full url check if a response is blank" do
-    expect(described_class.new({}).legal_urls?("", [post_1.id], [topic_1.id])).to eq(true)
+    expect(described_class.new({}).legal_post_urls?("", [post_1.id], [topic_1.id])).to eq(true)
   end
 
   it "correctly identifies an illegal topic id in a url in a response" do
-    expect(described_class.new({}).legal_urls?("hello /t/slug/113/2", [post_1.id], [topic_1.id])).to eq(false)
+    expect(described_class.new({}).legal_post_urls?("hello /t/slug/113/2", [post_1.id], [topic_1.id])).to eq(false)
+  end
+
+  it "correctly identifies an illegal non-post url in a response" do
+    expect(described_class.new({}).legal_non_post_urls?("hello https://someplace.com/t/slug/113/2 try looking at https://notanexample.com it's great", ["https://example.com", "https://otherexample.com"])).to eq(false)
+  end
+
+  it "correctly identifies a legal non-post url in a response" do
+    expect(described_class.new({}).legal_non_post_urls?("hello https://someplace.com/t/slug/113/2 try looking at https://example.com it's great", ["https://example.com", "https://otherexample.com"])).to eq(true)
   end
 end

spec/lib/functions/forum_search_function_spec.rb

@@ -42,7 +42,7 @@
     expect(topic_1).not_to be_nil
     expect(topic_2).not_to be_nil
     expect(topic_3).not_to be_nil
-    expect(subject.process(args)[:topic_ids_found]).to eq([post_2.topic_id])
+    expect(subject.process(args)[:topic_ids_found]).to eq([post_3.topic_id, post_5.topic_id])
     expect(subject.process(args)[:post_ids_found]).to include(post_5.id)
     expect(subject.process(args)[:post_ids_found]).to include(post_3.id)
     expect(subject.process(args)[:post_ids_found]).to include(post_2.id)
@@ -76,4 +76,19 @@
   it "finds urls with a post id" do
     expect(subject.find_post_and_topic_ids_from_raw_urls(post_5.raw)).to eq([[post_2.topic_id], [post_2.id]])
   end
-end
+
+  it "method finds urls that are not posts" do
+    text = <<~TEXT
+      Check out these links:
+      this one is good https://meta.discourse.org/t/user-specific-slow-mode/310081/1
+        wow https://example.com yeah
+      https://meta.discourse.org/t/another-topic/310082
+      http://another-example.org/path?query=string  this is a good one
+      my my https://meta.discourse.org/t/yet-another-topic/310083/2
+      (https://www.sample.org/sample-path)
+    TEXT
+    result = subject.find_other_urls(text)
+    expect(result.length).to eq(3)
+    expect(result).to eq(["https://example.com", "http://another-example.org/path?query=string", "https://www.sample.org/sample-path"])
+  end
+end