FIX: Do not return AI content if not in <output> (#223)

There is a bug now where if the new DiscourseAi translator does not return values back in `<translator>` tags, we return the full response.

This commit scopes both the input and output in its respective tags, and returns an empty string if there is no output tag. 

The problem only exists on topic translations, not post translations, as posts contain html and the prompt for translating indicates there will be html (topics do not have them). So an accompanying migration has been added

example incorrect result:
```
"To provide a translation, I first need the original text in a specific language to translate it into \"pt\". Could you please provide the text you wish to translate, including any HTML tags?"
```

Author: nattsw

app/services/discourse_ai/translator.rb

@@ -3,11 +3,13 @@
 module DiscourseAi
   class Translator
     PROMPT_TEMPLATE = <<~TEXT.freeze
-      You are a highly skilled linguist and web programmer, with expertise in many languages, and very well versed in HTML.
+      You are a highly skilled linguist of many languages and have expert knowledge in HTML.
       Your task is to identify the language of the text I provide and accurately translate it into this language locale "%{target_language}" while preserving the meaning, tone, and nuance of the original text.
-      The text will contain html tags, which must absolutely be preserved in the translation.
+      The text may or may not contain html tags. If they do, preserve them.
       Maintain proper grammar, spelling, and punctuation in the translated version.
-      Wrap the translated text in a <translation> tag.
+      You will find the text between <input></input> XML tags.
+      Include your translation between <output></output> XML tags.
+      Do not write explanations.
     TEXT
 
     def initialize(text, target_language)
@@ -19,7 +21,7 @@ def translate
       prompt =
         DiscourseAi::Completions::Prompt.new(
           build_prompt(@target_language),
-          messages: [{ type: :user, content: @text, id: "user" }],
+          messages: [{ type: :user, content: "<input>#{@text}</input>", id: "user" }],
         )
 
       llm_translation =
@@ -29,7 +31,7 @@ def translate
           feature_name: "translator-translate",
         )
 
-      (Nokogiri::HTML5.fragment(llm_translation).at("translation")&.inner_html || llm_translation)
+      (Nokogiri::HTML5.fragment(llm_translation).at("output")&.inner_html || "").strip
     end
 
     private

db/migrate/20250224120505_cleanup_ai_translations.rb

@@ -0,0 +1,16 @@
+# frozen_string_literal: true
+
+class CleanupAiTranslations < ActiveRecord::Migration[7.2]
+  def up
+    execute <<~SQL
+      DELETE FROM discourse_translator_topic_translations
+      WHERE translation LIKE 'To%'
+      AND translation ILIKE '%translat%'
+      AND LENGTH(translation) > 100;
+    SQL
+  end
+
+  def down
+    raise ActiveRecord::IrreversibleMigration
+  end
+end

spec/services/discourse_ai/translator_spec.rb

@@ -17,13 +17,15 @@
     it "creates the correct prompt" do
       allow(DiscourseAi::Completions::Prompt).to receive(:new).with(
         <<~TEXT,
-          You are a highly skilled linguist and web programmer, with expertise in many languages, and very well versed in HTML.
+          You are a highly skilled linguist of many languages and have expert knowledge in HTML.
           Your task is to identify the language of the text I provide and accurately translate it into this language locale "de" while preserving the meaning, tone, and nuance of the original text.
-          The text will contain html tags, which must absolutely be preserved in the translation.
+          The text may or may not contain html tags. If they do, preserve them.
           Maintain proper grammar, spelling, and punctuation in the translated version.
-          Wrap the translated text in a <translation> tag.
+          You will find the text between <input></input> XML tags.
+          Include your translation between <output></output> XML tags.
+          Do not write explanations.
         TEXT
-        messages: [{ type: :user, content: text_to_translate, id: "user" }],
+        messages: [{ type: :user, content: "<input>cats are great</input>", id: "user" }],
       ).and_call_original
 
       described_class.new(text_to_translate, target_language).translate
@@ -48,19 +50,17 @@
 
     it "returns the translation from the llm's response in the translation tag" do
       DiscourseAi::Completions::Llm.with_prepared_responses(
-        ["<translation>hur dur hur dur!</translation>"],
+        ["<output>hur dur hur dur!</output>"],
       ) do
         expect(
           described_class.new(text_to_translate, target_language).translate,
         ).to eq "hur dur hur dur!"
       end
     end
 
-    it "returns the raw response if the translation tag is not present" do
+    it "returns an empty string if the translation tag is not present" do
       DiscourseAi::Completions::Llm.with_prepared_responses(["raw response."]) do
-        expect(
-          described_class.new(text_to_translate, target_language).translate,
-        ).to eq "raw response."
+        expect(described_class.new(text_to_translate, target_language).translate).to eq ""
       end
     end
   end

spec/services/discourse_ai_spec.rb

@@ -42,7 +42,7 @@
 
     it "translates the post and returns [locale, translated_text]" do
       DiscourseAi::Completions::Llm.with_prepared_responses(
-        ["<translation>some translated text</translation>"],
+        ["<output>some translated text</output>"],
       ) do
         locale, translated_text = DiscourseTranslator::DiscourseAi.translate(post)
         expect(locale).to eq "de"