add system instruction

update models
introduce HarmCategoryType
update HarmCategory

Author: michael-ameri

gemini-api/src/main/java/swiss/ameri/gemini/api/GenAi.java

@@ -215,6 +215,8 @@ public Stream<GeneratedContent> generateContentStream(GenerativeModel model) {
                             var gcr = jsonParser.fromJson(line.substring(STREAM_LINE_PREFIX_LENGTH), GenerateContentResponse.class);
                             // each element can just replace the previous one
                             this.responseById.put(uuid, gcr);
+                            // todo catch safety error
+                            //if ("SAFETY".equals(gcr.candidates().get(0).finishReason())) {
                             return new GeneratedContent(uuid, gcr.candidates().get(0).content().parts().get(0).text());
                         } catch (Exception e) {
                             throw new RuntimeException("Unexpected line:\n" + line, e);
@@ -326,7 +328,18 @@ public CompletableFuture<List<ContentEmbedding>> embedContents(
 
     private static GenerateContentRequest convert(GenerativeModel model) {
         List<GenerationContent> generationContents = convertGenerationContents(model);
-        return new GenerateContentRequest(model.modelName(), generationContents, model.safetySettings(), model.generationConfig());
+        return new GenerateContentRequest(
+                model.modelName(),
+                generationContents,
+                model.safetySettings(),
+                model.generationConfig(),
+                model.systemInstruction().isEmpty() ? null :
+                        new SystemInstruction(
+                                model.systemInstruction().stream()
+                                        .map(SystemInstructionPart::new)
+                                        .toList()
+                        )
+        );
     }
 
     private static List<GenerationContent> convertGenerationContents(GenerativeModel model) {
@@ -578,7 +591,18 @@ private record GenerateContentRequest(
             String model,
             List<GenerationContent> contents,
             List<SafetySetting> safetySettings,
-            GenerationConfig generationConfig
+            GenerationConfig generationConfig,
+            SystemInstruction systemInstruction
+    ) {
+    }
+
+    private record SystemInstruction(
+            List<SystemInstructionPart> parts
+    ) {
+    }
+
+    private record SystemInstructionPart(
+            String text
     ) {
     }
 

gemini-api/src/main/java/swiss/ameri/gemini/api/GenerativeModel.java

@@ -6,16 +6,18 @@
 /**
  * Contains all the information needed for Gemini API to generate new content.
  *
- * @param modelName        to be used. see {@link ModelVariant}. Must start with "models/"
- * @param contents         given as input to Gemini API
- * @param safetySettings   optional, to adjust safety settings
- * @param generationConfig optional, to configure the prompt
+ * @param modelName         to be used. see {@link ModelVariant}. Must start with "models/"
+ * @param contents          given as input to Gemini API
+ * @param safetySettings    optional, to adjust safety settings
+ * @param generationConfig  optional, to configure the prompt
+ * @param systemInstruction optional, system instruction
  */
 public record GenerativeModel(
         String modelName,
         List<Content> contents,
         List<SafetySetting> safetySettings,
-        GenerationConfig generationConfig
+        GenerationConfig generationConfig,
+        List<String> systemInstruction
 ) {
 
     /**
@@ -35,6 +37,7 @@ public static class GenerativeModelBuilder {
         private GenerationConfig generationConfig;
         private final List<Content> contents = new ArrayList<>();
         private final List<SafetySetting> safetySettings = new ArrayList<>();
+        private final List<String> systemInstructions = new ArrayList<>();
 
         private GenerativeModelBuilder() {
         }
@@ -71,6 +74,17 @@ public GenerativeModelBuilder addContent(Content content) {
             return this;
         }
 
+        /**
+         * Add system instruction
+         *
+         * @param systemInstruction to be added
+         * @return this
+         */
+        public GenerativeModelBuilder addSystemInstruction(String systemInstruction) {
+            this.systemInstructions.add(systemInstruction);
+            return this;
+        }
+
         /**
          * Add safety setting
          *
@@ -99,7 +113,7 @@ public GenerativeModelBuilder generationConfig(GenerationConfig generationConfig
          * @return a completed (not necessarily validated) {@link GenerativeModel}
          */
         public GenerativeModel build() {
-            return new GenerativeModel(modelName, contents, safetySettings, generationConfig);
+            return new GenerativeModel(modelName, contents, safetySettings, generationConfig, systemInstructions);
         }
     }
 

gemini-api/src/main/java/swiss/ameri/gemini/api/ModelVariant.java

@@ -15,17 +15,21 @@ public enum ModelVariant {
      */
     GEMINI_1_5_FLASH("gemini-1.5-flash"),
     /**
-     * Natural language tasks, multi-turn text and code chat, and code generation.
+     * High volume and lower intelligence tasks.
      */
-    GEMINI_1_0_PRO("gemini-1.0-pro"),
+    GEMINI_1_5_FLASH_8B("gemini-1.5-flash-8b"),
     /**
-     * Visual-related tasks, like generating image descriptions or identifying objects in images.
+     * Natural language tasks, multi-turn text and code chat, and code generation.
      */
-    GEMINI_1_0_PRO_VISION("gemini-pro-vision"),
+    GEMINI_1_0_PRO("gemini-1.0-pro"),
     /**
      * Measuring the relatedness of text strings.
      */
     TEXT_EMBEDDING_004("text-embedding-004"),
+    /**
+     * Providing source-grounded answers to questions.
+     */
+    AQA("aqa"),
     ;
 
     private final String variant;

gemini-api/src/main/java/swiss/ameri/gemini/api/SafetySetting.java

@@ -1,5 +1,10 @@
 package swiss.ameri.gemini.api;
 
+import java.util.Arrays;
+import java.util.List;
+
+import static swiss.ameri.gemini.api.SafetySetting.HarmCategoryType.*;
+
 /**
  * Safety settings according to <a href="https://ai.google.dev/api/rest/v1beta/SafetySetting#harmblockthreshold">SafetySetting</a>.
  *
@@ -29,55 +34,81 @@ public static SafetySetting of(
         );
     }
 
+    public enum HarmCategoryType {
+        GEMINI,
+        PALM,
+        UNKNOWN
+    }
+
     /**
      * According to <a href="https://ai.google.dev/api/rest/v1beta/HarmCategory">HarmCategory</a>.
-     * Currently, only the first 4 seem to be recognized as input.
+     * See {@link #harmCategoryType} for which can be used as input to a model.
      */
     public enum HarmCategory {
+
         /**
          * Harasment content.
          */
-        HARM_CATEGORY_HARASSMENT,
+        HARM_CATEGORY_HARASSMENT(GEMINI),
         /**
          * Hate speech and content.
          */
-        HARM_CATEGORY_HATE_SPEECH,
+        HARM_CATEGORY_HATE_SPEECH(GEMINI),
         /**
          * Sexually explicit content.
          */
-        HARM_CATEGORY_SEXUALLY_EXPLICIT,
+        HARM_CATEGORY_SEXUALLY_EXPLICIT(GEMINI),
         /**
          * Dangerous content.
          */
-        HARM_CATEGORY_DANGEROUS_CONTENT,
+        HARM_CATEGORY_DANGEROUS_CONTENT(GEMINI),
+        /**
+         * Content that may be used to harm civic integrity.
+         */
+        HARM_CATEGORY_CIVIC_INTEGRITY(GEMINI),
         /**
          * Category is unspecified.
          */
-        HARM_CATEGORY_UNSPECIFIED,
+        HARM_CATEGORY_UNSPECIFIED(UNKNOWN),
         /**
          * Negative or harmful comments targeting identity and/or protected attribute.
          */
-        HARM_CATEGORY_DEROGATORY,
+        HARM_CATEGORY_DEROGATORY(PALM),
         /**
          * Content that is rude, disrespectful, or profane.
          */
-        HARM_CATEGORY_TOXICITY,
+        HARM_CATEGORY_TOXICITY(PALM),
         /**
          * Describes scenarios depicting violence against an individual or group, or general descriptions of gore.
          */
-        HARM_CATEGORY_VIOLENCE,
+        HARM_CATEGORY_VIOLENCE(PALM),
         /**
          * Contains references to sexual acts or other lewd content.
          */
-        HARM_CATEGORY_SEXUAL,
+        HARM_CATEGORY_SEXUAL(PALM),
         /**
          * Promotes unchecked medical advice.
          */
-        HARM_CATEGORY_MEDICAL,
+        HARM_CATEGORY_MEDICAL(PALM),
         /**
          * Dangerous content that promotes, facilitates, or encourages harmful acts.
          */
-        HARM_CATEGORY_DANGEROUS,
+        HARM_CATEGORY_DANGEROUS(PALM);
+        private final HarmCategoryType harmCategoryType;
+
+        HarmCategory(HarmCategoryType harmCategoryType) {
+            this.harmCategoryType = harmCategoryType;
+        }
+
+        public HarmCategoryType harmCategoryType() {
+            return harmCategoryType;
+        }
+
+        public static List<HarmCategory> harmCategoriesFor(HarmCategoryType type) {
+            return Arrays.stream(values())
+                    .filter(category -> category.harmCategoryType == type)
+                    .toList();
+        }
     }
 
     /**

gemini-tester/src/main/java/swiss/ameri/gemini/tester/GeminiTester.java

@@ -172,7 +172,7 @@ private static void listModels(GenAi genAi) {
     private static void textAndImage(GenAi genAi) throws IOException {
         System.out.println("----- text and image");
         var model = GenerativeModel.builder()
-                .modelName(ModelVariant.GEMINI_1_0_PRO_VISION)
+                .modelName(ModelVariant.GEMINI_1_5_FLASH)
                 .addContent(
                         Content.textAndMediaContentBuilder()
                                 .role(Content.Role.USER)