Emit no gen_ai.* attributes when LLM Observability is disabled

The attributes are derived from the LLMObs tag set, which the openai
decorators do not build with LLM Observability off. Drop the fallback that
resolved operation, model and ml_app from APM tags instead, and collapse
GenAiApmTags back to a single apply(span).

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Author: mz1119

dd-java-agent/instrumentation/openai-java/openai-java-3.0/src/main/java/datadog/trace/instrumentation/openai_java/OpenAiDecorator.java

@@ -1,6 +1,5 @@
 package datadog.trace.instrumentation.openai_java;
 
-import static datadog.trace.api.llmobs.GenAiApmTags.stringTag;
 import static datadog.trace.bootstrap.instrumentation.api.AgentSpan.fromContext;
 
 import com.openai.core.ClientOptions;
@@ -18,7 +17,6 @@
 import datadog.trace.bootstrap.instrumentation.api.AgentSpanContext;
 import datadog.trace.bootstrap.instrumentation.api.AgentTracer;
 import datadog.trace.bootstrap.instrumentation.api.InternalSpanTypes;
-import datadog.trace.bootstrap.instrumentation.api.Tags;
 import datadog.trace.bootstrap.instrumentation.api.UTF8BytesString;
 import datadog.trace.bootstrap.instrumentation.decorator.ClientDecorator;
 import java.util.List;
@@ -40,8 +38,6 @@ public class OpenAiDecorator extends ClientDecorator {
   private static final String TOKENS_LIMIT_METRIC = METRIC_PREFIX + "tokens.limit";
   private static final String TOKENS_REMAINING_METRIC = METRIC_PREFIX + "tokens.remaining";
 
-  private static final String EMBEDDINGS_ENDPOINT = "/v1/embeddings";
-
   private static final String HEADER_PREFIX = "x-ratelimit-";
   private static final String LIMIT_REQUESTS_HEADER = HEADER_PREFIX + "limit-requests";
   private static final String REMAINING_REQUESTS_HEADER = HEADER_PREFIX + "remaining-requests";
@@ -193,30 +189,10 @@ protected void doBeforeFinish(@Nonnull Context context) {
         LLMObsMetricCollector.get()
             .recordSpanFinished(INTEGRATION, spanKind, isRootSpan, true, span.isError(), false);
       }
-    } else if (span != null) {
-      // Tracing still runs with LLM Observability off, where these remain resolvable.
-      GenAiApmTags.apply(
-          span, operationName(span), requestedModel(span), Config.get().getLlmObsMlApp());
     }
     super.doBeforeFinish(context);
   }
 
-  private static String operationName(AgentSpan span) {
-    String endpoint = stringTag(span, CommonTags.OPENAI_REQUEST_ENDPOINT);
-    if (endpoint == null) {
-      return null;
-    }
-    return EMBEDDINGS_ENDPOINT.equals(endpoint)
-        ? Tags.LLMOBS_EMBEDDING_SPAN_KIND
-        : Tags.LLMOBS_LLM_SPAN_KIND;
-  }
-
-  /** The response model resolves aliases the request used, so it wins. */
-  private static String requestedModel(AgentSpan span) {
-    String model = stringTag(span, CommonTags.OPENAI_RESPONSE_MODEL);
-    return model != null ? model : stringTag(span, CommonTags.OPENAI_REQUEST_MODEL);
-  }
-
   public void withHttpResponse(AgentSpan span, Headers headers) {
     if (!llmObsEnabled) {
       return;

dd-java-agent/instrumentation/openai-java/openai-java-3.0/src/test/java/datadog/trace/instrumentation/openai_java/LlmObsContextPropagationForkedTest.java

@@ -3,6 +3,7 @@
 import static org.junit.jupiter.api.Assertions.assertEquals;
 import static org.junit.jupiter.api.Assertions.assertNotNull;
 import static org.junit.jupiter.api.Assertions.assertNull;
+import static org.junit.jupiter.api.Assertions.assertTrue;
 
 import com.openai.client.OpenAIClient;
 import com.openai.client.okhttp.OpenAIOkHttpClient;
@@ -329,51 +330,42 @@ void parentlessOpenAiRequestSpanIsDroppedAtZeroSampleRate() throws Exception {
 }
 
 /**
- * Verifies the gen_ai.* attributes an openai.request span carries with LLM Observability disabled:
- * operation, model, provider and application, but never token usage or conversation id.
+ * Verifies that an openai.request span carries no gen_ai.* attributes at all with LLM Observability
+ * disabled. The attributes are derived from the LLMObs tag set, which the decorator does not build
+ * in that configuration.
  */
 @WithConfig(key = "llmobs.enabled", value = "false")
 class LlmObsDisabledForkedTest extends AbstractLlmObsOpenAiForkedTest {
 
   @Test
-  void chatCompletionEmitsTheGenAiAttributesAvailableWithoutLlmObs() {
+  void chatCompletionEmitsNoGenAiAttributes() {
     try {
       openAiClient.chat().completions().create(buildMinimalChatParams());
     } catch (Exception ignored) {
       // The mock server returns no body, so the SDK may throw while parsing the response.
     }
 
-    DDSpan openAiSpan = awaitOpenAiSpan("/v1/chat/completions");
-
-    assertEquals("llm", openAiSpan.getTag("gen_ai.operation.name"));
-    // The mock returns no body, so the request model stands in for the absent response model.
-    assertEquals(
-        openAiSpan.getTag("openai.request.model"), openAiSpan.getTag("gen_ai.request.model"));
-    assertEquals("openai", openAiSpan.getTag("gen_ai.provider.name"));
-    assertNotNull(openAiSpan.getTag("gen_ai.application.name"));
-
-    assertNull(openAiSpan.getTag("gen_ai.conversation.id"));
-    assertNull(openAiSpan.getTag("gen_ai.usage.input_tokens"));
-    assertNull(openAiSpan.getTag("gen_ai.usage.output_tokens"));
-    assertNull(openAiSpan.getTag("gen_ai.usage.total_tokens"));
-
-    assertNull(openAiSpan.getTag("_ml_obs_tag.span.kind"));
+    assertNoGenAiTags(awaitOpenAiSpan("/v1/chat/completions"));
   }
 
   @Test
-  void embeddingMapsToTheEmbeddingOperation() {
+  void embeddingEmitsNoGenAiAttributes() {
     try {
       openAiClient.embeddings().create(buildMinimalEmbeddingParams());
     } catch (Exception ignored) {
       // The mock server returns no body, so the SDK may throw while parsing the response.
     }
 
-    DDSpan openAiSpan = awaitOpenAiSpan("/v1/embeddings");
+    assertNoGenAiTags(awaitOpenAiSpan("/v1/embeddings"));
+  }
 
-    assertEquals("embedding", openAiSpan.getTag("gen_ai.operation.name"));
-    assertEquals(
-        openAiSpan.getTag("openai.request.model"), openAiSpan.getTag("gen_ai.request.model"));
-    assertEquals("openai", openAiSpan.getTag("gen_ai.provider.name"));
+  private static void assertNoGenAiTags(DDSpan span) {
+    // The endpoint and model the span was selected by are APM tags of their own, still present.
+    assertNotNull(span.getTag("openai.request.model"));
+    assertTrue(
+        span.getTags().keySet().stream().noneMatch(key -> key.startsWith("gen_ai.")),
+        "openai.request span should carry no gen_ai.* tags");
+    assertNull(span.getTag("_ml_obs_tag.span.kind"));
   }
 
   // Both tests here produce an openai.request span, so match on the endpoint rather than take the

internal-api/src/main/java/datadog/trace/api/llmobs/GenAiApmTags.java

@@ -46,21 +46,16 @@ public final class GenAiApmTags {
     {LLMOBS_METRIC_PREFIX + "reasoning_output_tokens", USAGE_REASONING_OUTPUT_TOKENS},
   };
 
-  public static void apply(AgentSpan span) {
-    apply(span, null, null, null);
-  }
-
   /**
    * Writes the attributes onto a span that is not yet finished, reading them from its {@code
-   * _ml_obs_tag.} / {@code _ml_obs_metric.} tags. The arguments take precedence over those tags and
-   * cover instrumentation that traces with LLM Observability disabled, where they are not all set.
-   * No-op for a span with no resolvable operation.
+   * _ml_obs_tag.} / {@code _ml_obs_metric.} tags. No-op for a span with no LLM Observability span
+   * kind.
    */
-  public static void apply(AgentSpan span, String operationName, String modelName, String mlApp) {
+  public static void apply(AgentSpan span) {
     if (span == null) {
       return;
     }
-    String operation = firstNonEmpty(operationName, stringTag(span, SPAN_KIND_TAG));
+    String operation = stringTag(span, SPAN_KIND_TAG);
     if (operation == null) {
       return;
     }
@@ -70,7 +65,7 @@ public static void apply(AgentSpan span, String operationName, String modelName,
         Tags.LLMOBS_LLM_SPAN_KIND.equals(operation)
             || Tags.LLMOBS_EMBEDDING_SPAN_KIND.equals(operation);
 
-    String model = firstNonEmpty(modelName, stringTag(span, MODEL_NAME_TAG));
+    String model = stringTag(span, MODEL_NAME_TAG);
     if (model != null || modelBacked) {
       span.setTag(REQUEST_MODEL, model == null ? DEFAULT_MODEL : model);
     }
@@ -79,7 +74,7 @@ public static void apply(AgentSpan span, String operationName, String modelName,
       span.setTag(
           PROVIDER_NAME, (provider == null ? DEFAULT_MODEL : provider).toLowerCase(Locale.ROOT));
     }
-    String application = firstNonEmpty(mlApp, stringTag(span, ML_APP_TAG));
+    String application = stringTag(span, ML_APP_TAG);
     if (application != null) {
       span.setTag(APPLICATION_NAME, application);
     }
@@ -100,7 +95,7 @@ public static void apply(AgentSpan span, String operationName, String modelName,
   }
 
   /** The value of {@code key} as a non-empty string, or null. */
-  public static String stringTag(AgentSpan span, String key) {
+  private static String stringTag(AgentSpan span, String key) {
     Object value = span.getTag(key);
     if (value == null) {
       return null;
@@ -109,9 +104,5 @@ public static String stringTag(AgentSpan span, String key) {
     return string.isEmpty() ? null : string;
   }
 
-  private static String firstNonEmpty(String preferred, String fallback) {
-    return preferred == null || preferred.isEmpty() ? fallback : preferred;
-  }
-
   private GenAiApmTags() {}
 }

internal-api/src/test/java/datadog/trace/api/llmobs/GenAiApmTagsTest.java

@@ -152,49 +152,6 @@ void spanWithoutLlmObsKindEmitsNothing() {
   @Test
   void nullSpanIsANoOp() {
     GenAiApmTags.apply(null);
-    GenAiApmTags.apply(null, Tags.LLMOBS_LLM_SPAN_KIND, "gpt-4", "app");
-  }
-
-  @Test
-  void argumentsStandInForMissingLlmObsTags() {
-    llmObsTag(LLMObsTags.MODEL_PROVIDER, "OpenAI");
-
-    GenAiApmTags.apply(span, Tags.LLMOBS_LLM_SPAN_KIND, "gpt-4", "my-app");
-
-    assertEquals(Tags.LLMOBS_LLM_SPAN_KIND, tags.get(GenAiApmTags.OPERATION_NAME));
-    assertEquals("gpt-4", tags.get(GenAiApmTags.REQUEST_MODEL));
-    assertEquals("openai", tags.get(GenAiApmTags.PROVIDER_NAME));
-    assertEquals("my-app", tags.get(GenAiApmTags.APPLICATION_NAME));
-    assertFalse(tags.containsKey(GenAiApmTags.CONVERSATION_ID));
-    assertFalse(tags.containsKey(GenAiApmTags.USAGE_INPUT_TOKENS));
-  }
-
-  @Test
-  void argumentsTakePrecedenceOverLlmObsTags() {
-    llmObsTag(Tags.SPAN_KIND, Tags.LLMOBS_WORKFLOW_SPAN_KIND);
-    llmObsTag(LLMObsTags.MODEL_NAME, "gpt-3.5");
-    llmObsTag(LLMObsTags.ML_APP, "tagged-app");
-
-    GenAiApmTags.apply(span, Tags.LLMOBS_LLM_SPAN_KIND, "gpt-4", "my-app");
-
-    assertEquals(Tags.LLMOBS_LLM_SPAN_KIND, tags.get(GenAiApmTags.OPERATION_NAME));
-    assertEquals("gpt-4", tags.get(GenAiApmTags.REQUEST_MODEL));
-    assertEquals("my-app", tags.get(GenAiApmTags.APPLICATION_NAME));
-  }
-
-  @Test
-  void modelBackedArgumentsFallBackToCustom() {
-    GenAiApmTags.apply(span, Tags.LLMOBS_EMBEDDING_SPAN_KIND, null, "app");
-
-    assertEquals("custom", tags.get(GenAiApmTags.REQUEST_MODEL));
-    assertEquals("custom", tags.get(GenAiApmTags.PROVIDER_NAME));
-  }
-
-  @Test
-  void spanWithNoResolvableOperationEmitsNothing() {
-    GenAiApmTags.apply(span, null, "gpt-4", "my-app");
-
-    assertTrue(tags.keySet().stream().noneMatch(key -> key.startsWith("gen_ai.")));
   }
 
   private void llmObsTag(String key, String value) {