Optimize context prompt

Author: pbernet

src/main/scala/tools/AnthropicCompletions.java

@@ -20,7 +20,11 @@
  */
 public class AnthropicCompletions {
     private static final Logger LOGGER = LoggerFactory.getLogger(AnthropicCompletions.class);
+    // Add your API key
     public static final String API_KEY = "***";
+
+    // Note that sonnet 4 has a different output style
+    // https://docs.anthropic.com/en/docs/about-claude/models/overview#model-comparison-table
     public static final String MODEL_NAME = "claude-3-7-sonnet-20250219";
 
     private final ChatModel model;
@@ -79,12 +83,12 @@ public ImmutablePair<String, Integer> runCompletions(String prompt) {
     }
 
     public static void main(String[] args) {
-        String toTranslate = CompletionsUtil.createTranslationPrompt("This is fun.", "English", "German");
+        String prompt = CompletionsUtil.createTranslationPrompt("This is fun.", "English", "German");
 
-        ImmutablePair<String, Integer> result = new AnthropicCompletions().runCompletions(toTranslate);
+        ImmutablePair<String, Integer> result = new AnthropicCompletions().runCompletions(prompt);
         CompletionsUtil.logCompletionResult(result.getLeft(), result.getRight(), "Translation");
 
-        ImmutablePair<String, Integer> resultWithContext = AnthropicCompletions.withContext("The Hangover", 2009).runCompletions(toTranslate);
+        ImmutablePair<String, Integer> resultWithContext = AnthropicCompletions.withContext("The Hangover", 2009).runCompletions(prompt);
         CompletionsUtil.logCompletionResult(resultWithContext.getLeft(), resultWithContext.getRight(), "Translation with context");
     }
 

src/main/scala/tools/CompletionsUtil.java

@@ -28,22 +28,22 @@ public static String createTranslationPrompt(String text, String sourceLanguage,
     }
 
     /**
-     * Creates a movie context prompt
+     * Creates a movie context prompt optimized for subtitle translation.
+     * Focuses on elements most relevant to understanding dialogue and cultural context.
      *
      * @param movieTitle       The title of the movie
      * @param movieReleaseYear The release year of the movie
-     * @return Formatted movie context prompt
+     * @return Formatted movie context prompt optimized for subtitle translation
      */
     public static String createMovieContextPrompt(String movieTitle, int movieReleaseYear) {
         return String.format(
-                "Find movie metadata about the movie '%s' released in %d\n" +
-                        "If you don't know this movie or if the title or release year is ambiguous: Respond with \"N/A\", do not apologize." +
-                        "Structure of the response:\n" +
-                        "Plot summary: [Max 100 words]" +
-                        "Locations: [Max 3 key locations]" +
-                        "Key characters: [Max 5 main character names, no actor names]" +
-                        "Themes: [Max 3 main themes]" +
-                        "Notable scenes: [Max 3 brief descriptions of memorable scenes]",
+                "Provide context for subtitle translation of '%s' (%d).\n" +
+                        "If unknown/ambiguous: respond \"N/A\".\n" +
+                        "Genre: [1-2 genres]\n" +
+                        "Setting: [Time period, main location]\n" +
+                        "Key characters: [Max 4 main character names with brief role]\n" +
+                        "Cultural context: [Important cultural/historical references for translation, Max 100 words]\n" +
+                        "Language style: [Formal/informal, period-specific terms, slang]",
                 movieTitle, movieReleaseYear);
     }
 

src/main/scala/tools/OpenAICompletions.java

@@ -22,6 +22,7 @@ public class OpenAICompletions {
     private static final Logger LOGGER = LoggerFactory.getLogger(OpenAICompletions.class);
     // Add your API key, see: https://platform.openai.com/api-keys
     public static final String API_KEY = "***";
+
     public static final String MODEL_NAME = "gpt-4o";
 
     private final ChatModel model;
@@ -78,12 +79,12 @@ public ImmutablePair<String, Integer> runCompletions(String prompt) {
     }
 
     public static void main(String[] args) {
-        String toTranslate = CompletionsUtil.createTranslationPrompt("This is fun.", "English", "German");
+        String prompt = CompletionsUtil.createTranslationPrompt("This is fun.", "English", "German");
 
-        ImmutablePair<String, Integer> result = new OpenAICompletions().runCompletions(toTranslate);
+        ImmutablePair<String, Integer> result = new OpenAICompletions().runCompletions(prompt);
         CompletionsUtil.logCompletionResult(result.getLeft(), result.getRight(), "Translation");
 
-        ImmutablePair<String, Integer> resultWithContext = OpenAICompletions.withContext("The Hangover", 2009).runCompletions(toTranslate);
+        ImmutablePair<String, Integer> resultWithContext = OpenAICompletions.withContext("The Hangover", 2009).runCompletions(prompt);
         CompletionsUtil.logCompletionResult(resultWithContext.getLeft(), resultWithContext.getRight(), "Translation with context");
     }
 

src/main/scala/tools/SubtitleTranslator.scala

@@ -21,13 +21,13 @@ import scala.util.{Failure, Success}
   *  - Continuously write translated blocks to target file
   *    *
   * Usage:
-  *  - Wire Params
-  *  - Add API_KEY in [[AnthropicCompletions]], [[OpenAICompletions]]  and then run this class
+  *  - Wire Params, eg sourceFilePath
+  *  - Decide, whether to use context info to streamline translation
+  *  - Add API_KEY in [[AnthropicCompletions]], [[OpenAICompletions]] and then run this class
   *  - Scan log for WARN log messages and improve corresponding blocks in target file manually
   *
   * Remarks:
   *  - Numerical block headers in the .srt files are not interpreted, only timestamps matter
-  *    . - Default params are just example values
   *
   * Similar to: [[sample.stream.SessionWindow]]
   */
@@ -37,14 +37,15 @@ object SubtitleTranslator extends App {
   implicit val executionContext: ExecutionContextExecutor = system.dispatcher
 
   // Params
-  private val sourceFilePath = "Killers.Of.The.Flower.Moon.2023.720p.WEBRip.x264.AAC-[YTS.MX].srt"
-  private val targetFilePath = "DE_NEW_Killers.Of.The.Flower.Moon.2023.720p.WEBRip.x264.AAC-[YTS.MX].srt"
+  private val sourceFilePath = "EN_Killers.Of.The.Flower.Moon.srt"
+  private val targetFilePath = "DE_Killers.Of.The.Flower.Moon.srt"
   private val targetLanguage = "German"
   private val movieTitle = "Killers of the Flower Moon"
   private val movieReleaseYear = 2023
 
-  private val defaultModel = OpenAICompletions.withContext(movieTitle, movieReleaseYear)
-  private val fallbackModel = AnthropicCompletions.withContext(movieTitle, movieReleaseYear)
+  private val useContext = true
+  private val defaultModel = if (useContext) OpenAICompletions.withContext(movieTitle, movieReleaseYear) else new OpenAICompletions()
+  private val fallbackModel = if (useContext) AnthropicCompletions.withContext(movieTitle, movieReleaseYear) else new AnthropicCompletions()
 
   private val maxGapSeconds = 1 // gap time between two scenes (= session windows)
   private val endLineTag = "\n"