feat: support protected_material_code as an output content filtering module

docs/release_notes.md

@@ -18,6 +18,7 @@
 
 - [Orchestration] Added new API `OrchestrationTemplateReference#withScope` to support prompt templates with resource-group scope.
 - [Orchestration] Chat completion calls now can have multiple module configs to support [fallback modules](https://sap.github.io/ai-sdk/docs/java/orchestration/chat-completion).
+- [Orchestration] `protected_material_code` is now supported as an output content filtering module .
 
 ### 🐛 Fixed Issues
 

orchestration/src/main/java/com/sap/ai/sdk/orchestration/AzureContentFilter.java

@@ -48,6 +48,9 @@ public class AzureContentFilter implements ContentFilter {
   /** The filter category for violence content. */
   @Nullable AzureFilterThreshold violence;
 
+  /** The filter category for protected material content. */
+  boolean protectedMaterialCode;
+
   /**
    * A flag to set prompt shield on input filer.
    *
@@ -90,7 +93,11 @@ public AzureContentSafetyInputFilterConfig createInputFilterConfig() {
   @Override
   @Nonnull
   public AzureContentSafetyOutputFilterConfig createOutputFilterConfig() {
-    if (hate == null && selfHarm == null && sexual == null && violence == null) {
+    if (hate == null
+        && selfHarm == null
+        && sexual == null
+        && violence == null
+        && !protectedMaterialCode) {
       throw new IllegalArgumentException("At least one filter category must be set");
     }
 
@@ -101,6 +108,7 @@ public AzureContentSafetyOutputFilterConfig createOutputFilterConfig() {
                 .hate(hate != null ? hate.getAzureThreshold() : null)
                 .selfHarm(selfHarm != null ? selfHarm.getAzureThreshold() : null)
                 .sexual(sexual != null ? sexual.getAzureThreshold() : null)
-                .violence(violence != null ? violence.getAzureThreshold() : null));
+                .violence(violence != null ? violence.getAzureThreshold() : null)
+                .protectedMaterialCode(true));
   }
 }

orchestration/src/test/java/com/sap/ai/sdk/orchestration/OrchestrationModuleConfigTest.java

@@ -50,18 +50,21 @@ static class InsideTestClass {
   void testStackingInputAndOutputFilter() {
     final var config = new OrchestrationModuleConfig().withLlmConfig(GPT_4O);
 
-    final var filter =
+    final var inputFilter =
         new AzureContentFilter()
             .hate(ALLOW_SAFE_LOW_MEDIUM)
             .selfHarm(ALLOW_SAFE_LOW_MEDIUM)
             .sexual(ALLOW_SAFE_LOW_MEDIUM)
             .violence(ALLOW_SAFE_LOW_MEDIUM);
+    final var outputFilter = inputFilter.protectedMaterialCode(true);
 
-    final var configWithInputFirst = config.withInputFiltering(filter).withOutputFiltering(filter);
+    final var configWithInputFirst =
+        config.withInputFiltering(inputFilter).withOutputFiltering(outputFilter);
     assertThat(configWithInputFirst.getFilteringConfig()).isNotNull();
     assertThat(configWithInputFirst.getFilteringConfig().getInput()).isNotNull();
 
-    final var configWithOutputFirst = config.withOutputFiltering(filter).withInputFiltering(filter);
+    final var configWithOutputFirst =
+        config.withOutputFiltering(outputFilter).withInputFiltering(inputFilter);
     assertThat(configWithOutputFirst.getFilteringConfig()).isNotNull();
     assertThat(configWithOutputFirst.getFilteringConfig().getOutput()).isNotNull();
   }

orchestration/src/test/java/com/sap/ai/sdk/orchestration/OrchestrationUnitTest.java

@@ -438,18 +438,21 @@ void filteringLoose() throws IOException {
                     .withBodyFile("filteringLooseResponse.json")
                     .withHeader("Content-Type", "application/json")));
 
-    final var azureFilter =
+    final var azureInputFilter =
         new AzureContentFilter()
             .hate(ALLOW_SAFE_LOW_MEDIUM)
             .selfHarm(ALLOW_SAFE_LOW_MEDIUM)
             .sexual(ALLOW_SAFE_LOW_MEDIUM)
             .violence(ALLOW_SAFE_LOW_MEDIUM);
+    final var azureOutputFilter = azureInputFilter.protectedMaterialCode(true);
 
     final var llamaFilter = new LlamaGuardFilter().config(LlamaGuard38b.create().selfHarm(true));
 
     client.chatCompletion(
         prompt,
-        config.withInputFiltering(azureFilter, llamaFilter).withOutputFiltering(azureFilter));
+        config
+            .withInputFiltering(azureInputFilter, llamaFilter)
+            .withOutputFiltering(azureOutputFilter));
     // the result is asserted in the verify step below
 
     // verify that null fields are absent from the sent request
@@ -464,19 +467,20 @@ void filteringLooseStream() throws IOException {
         post(anyUrl())
             .willReturn(aResponse().withBody(res).withHeader("Content-Type", "application/json")));
 
-    final var azureFilter =
+    final var azureInputFilter =
         new AzureContentFilter()
             .hate(ALLOW_SAFE_LOW_MEDIUM)
             .selfHarm(ALLOW_SAFE_LOW_MEDIUM)
             .sexual(ALLOW_SAFE_LOW_MEDIUM)
             .violence(ALLOW_SAFE_LOW_MEDIUM);
+    final var azureOutputFilter = azureInputFilter.protectedMaterialCode(true);
 
     final var llamaFilter = new LlamaGuardFilter().config(LlamaGuard38b.create().selfHarm(true));
 
     OrchestrationModuleConfig myConfig =
         config
-            .withInputFiltering(azureFilter, llamaFilter)
-            .withOutputFiltering(azureFilter)
+            .withInputFiltering(azureInputFilter, llamaFilter)
+            .withOutputFiltering(azureOutputFilter)
             .withOutputFilteringStreamOptions(FilteringStreamOptions.create().overlap(1_000));
 
     Stream<String> result = client.streamChatCompletion(prompt, myConfig);
@@ -580,7 +584,8 @@ void outputFilteringStrict() {
             .hate(ALLOW_SAFE)
             .selfHarm(ALLOW_SAFE)
             .sexual(ALLOW_SAFE)
-            .violence(ALLOW_SAFE);
+            .violence(ALLOW_SAFE)
+            .protectedMaterialCode(true);
 
     final var llamaFilter =
         new LlamaGuardFilter().config(LlamaGuard38b.create().violentCrimes(true));
@@ -601,7 +606,8 @@ void outputFilteringStrict() {
                               "Hate", 6,
                               "SelfHarm", 0,
                               "Sexual", 0,
-                              "Violence", 6),
+                              "Violence", 6,
+                              "ProtectedMaterialCode", false),
                           "llama_guard_3_8b",
                           Map.of("violent_crimes", true)));
               assertThat(e.getErrorResponse()).isNull();
@@ -612,6 +618,7 @@ void outputFilteringStrict() {
               assertThat(e.getAzureContentSafetyOutput().getSelfHarm()).isEqualTo(NUMBER_0);
               assertThat(e.getAzureContentSafetyOutput().getSexual()).isEqualTo(NUMBER_0);
               assertThat(e.getAzureContentSafetyOutput().getViolence()).isEqualTo(NUMBER_6);
+              assertThat(e.getAzureContentSafetyOutput().isProtectedMaterialCode()).isFalse();
 
               assertThat(e.getLlamaGuard38b()).isNotNull();
               assertThat(e.getLlamaGuard38b().isViolentCrimes()).isTrue();

orchestration/src/test/resources/__files/outputFilteringStrict.json

@@ -21,7 +21,8 @@
               "Hate": 6,
               "SelfHarm": 0,
               "Sexual": 0,
-              "Violence": 6
+              "Violence": 6,
+              "ProtectedMaterialCode": false
             },
             "llama_guard_3_8b": {
               "violent_crimes": true

orchestration/src/test/resources/filteringLooseRequestStream.json

@@ -56,7 +56,7 @@
                 "self_harm": 4,
                 "sexual": 4,
                 "violence": 4,
-                "protected_material_code" : false
+                "protected_material_code" : true
               }
             }
           ],

sample-code/spring-app/src/main/java/com/sap/ai/sdk/app/services/OrchestrationService.java

@@ -204,6 +204,7 @@ public OrchestrationChatResponse inputFiltering(@Nonnull final AzureFilterThresh
    */
   @Nonnull
   public OrchestrationChatResponse outputFiltering(@Nonnull final AzureFilterThreshold policy) {
+    val isProtected = true; // enforce protected material code filtering
 
     val systemMessage = Message.system("Give three paraphrases for the following sentence");
     // Reliably triggering the content filter of models fine-tuned for ethical compliance
@@ -212,7 +213,12 @@ public OrchestrationChatResponse outputFiltering(@Nonnull final AzureFilterThres
         new OrchestrationPrompt("'We shall spill blood tonight', said the operation in-charge.")
             .messageHistory(List.of(systemMessage));
     val filterConfig =
-        new AzureContentFilter().hate(policy).selfHarm(policy).sexual(policy).violence(policy);
+        new AzureContentFilter()
+            .hate(policy)
+            .selfHarm(policy)
+            .sexual(policy)
+            .violence(policy)
+            .protectedMaterialCode(isProtected);
 
     val configWithFilter = config.withOutputFiltering(filterConfig);
     return client.chatCompletion(prompt, configWithFilter);

sample-code/spring-app/src/test/java/com/sap/ai/sdk/app/controllers/OrchestrationTest.java

@@ -286,6 +286,7 @@ void testOutputFilteringStrict() {
               assertThat(actualAzureContentSafety.getSelfHarm()).isEqualTo(NUMBER_0);
               assertThat(actualAzureContentSafety.getSexual()).isEqualTo(NUMBER_0);
               assertThat(actualAzureContentSafety.getHate()).isEqualTo(NUMBER_0);
+              assertThat(actualAzureContentSafety.isProtectedMaterialCode()).isFalse();
             });
   }
 
@@ -299,7 +300,8 @@ void testOutputFilteringLenient() {
     assertThat(response.getContent()).isNotEmpty();
 
     var filterResult = response.getOriginalResponse().getIntermediateResults().getOutputFiltering();
-    assertThat(filterResult.getMessage()).containsPattern("Choice 0: Filtering was skipped.");
+    assertThat(filterResult.getMessage())
+        .containsPattern("Choice 0: Filtering passed successfully.");
   }
 
   @Test