Added GenAI support for Amazon Nova model (#153)

Co-authored-by: Michael He <[email protected]>

Author: ezhang6811

src/OpenTelemetry.Instrumentation.AWS/Implementation/AWSLlmModelProcessor.cs

@@ -37,7 +37,11 @@ internal static void ProcessGenAiAttributes<T>(Activity activity, T message, str
                     }
 
                     // extract model specific attributes based on model name
-                    if (modelName.Contains("amazon.titan"))
+                    if (modelName.Contains("amazon.nova"))
+                    {
+                        ProcessNovaModelAttributes(activity, jsonObject, isRequest);
+                    }
+                    else if (modelName.Contains("amazon.titan"))
                     {
                         ProcessTitanModelAttributes(activity, jsonObject, isRequest);
                     }
@@ -70,6 +74,55 @@ internal static void ProcessGenAiAttributes<T>(Activity activity, T message, str
         }
     }
 
+    private static void ProcessNovaModelAttributes(Activity activity, Dictionary<string, JsonElement> jsonBody, bool isRequest)
+    {
+        try
+        {
+            if (isRequest)
+            {
+                if (jsonBody.TryGetValue("inferenceConfig", out var inferenceConfig))
+                {
+                    if (inferenceConfig.TryGetProperty("top_p", out var topP))
+                    {
+                        activity.SetTag(AWSSemanticConventions.AttributeGenAiTopP, topP.GetDouble());
+                    }
+
+                    if (inferenceConfig.TryGetProperty("temperature", out var temperature))
+                    {
+                        activity.SetTag(AWSSemanticConventions.AttributeGenAiTemperature, temperature.GetDouble());
+                    }
+
+                    if (inferenceConfig.TryGetProperty("max_new_tokens", out var maxTokens))
+                    {
+                        activity.SetTag(AWSSemanticConventions.AttributeGenAiMaxTokens, maxTokens.GetInt32());
+                    }
+                }
+            }
+            else
+            {
+                if (jsonBody.TryGetValue("usage", out var usage))
+                {
+                    if (usage.TryGetProperty("inputTokens", out var inputTokens))
+                    {
+                        activity.SetTag(AWSSemanticConventions.AttributeGenAiInputTokens, inputTokens.GetInt32());
+                    }
+                    if (usage.TryGetProperty("outputTokens", out var outputTokens))
+                    {
+                        activity.SetTag(AWSSemanticConventions.AttributeGenAiOutputTokens, outputTokens.GetInt32());
+                    }
+                }
+                if (jsonBody.TryGetValue("stopReason", out var finishReasons))
+                {
+                    activity.SetTag(AWSSemanticConventions.AttributeGenAiFinishReasons, new string[] { finishReasons.GetString() ?? string.Empty });
+                }
+            }
+        }
+        catch (Exception ex)
+        {
+            Console.WriteLine("Exception: " + ex.Message);
+        }
+    }
+
     private static void ProcessTitanModelAttributes(Activity activity, Dictionary<string, JsonElement> jsonBody, bool isRequest)
     {
         try

test/contract-tests/images/applications/TestSimpleApp.AWSSDK.Core/BedrockTests.cs

@@ -39,9 +39,55 @@ public GetGuardrailResponse GetGuardrailResponse()
         };
     }
 
-    // 6 InvokeModel test calls and responses, one for each supported model
+    // 7 InvokeModel test calls and responses, one for each supported model
     // The manual responses are automatically serialized to a MemoryStream and used as the response body
 
+    public void InvokeModelAmazonNova()
+    {
+        bedrockRuntime.InvokeModelAsync(new InvokeModelRequest
+        {
+            ModelId = "us.amazon.nova-micro-v1:0",
+            Body = new MemoryStream(Encoding.UTF8.GetBytes(JsonSerializer.Serialize(new
+            {
+                messages = new object[]
+                {
+                    new
+                    {
+                        role = "user",
+                        content = new object[]
+                        {
+                            new
+                            {
+                                text = "sample input text",
+                            }
+                        }
+                    },
+                },
+                inferenceConfig = new
+                {
+                    temperature = 0.123,
+                    top_p = 0.456,
+                    max_new_tokens = 123,
+                },
+            }))),
+            ContentType = "application/json",
+        });
+        return;
+    }
+
+    public object InvokeModelAmazonNovaResponse()
+    {
+        return new
+        {
+            usage = new
+            {
+                inputTokens = 456,
+                outputTokens = 789,
+            },
+            stopReason = "finish_reason",
+        };
+    }
+
     public void InvokeModelAmazonTitan()
     {
         bedrockRuntime.InvokeModelAsync(new InvokeModelRequest

test/contract-tests/images/applications/TestSimpleApp.AWSSDK.Core/Program.cs

@@ -169,6 +169,10 @@
     .WithName("get-guardrail")
     .WithOpenApi();
 
+app.MapGet("bedrock/invokemodel/invoke-model-nova", (BedrockTests bedrock) => bedrock.InvokeModelAmazonNova())
+    .WithName("invoke-model-nova")
+    .WithOpenApi();
+
 app.MapGet("bedrock/invokemodel/invoke-model-titan", (BedrockTests bedrock) => bedrock.InvokeModelAmazonTitan())
     .WithName("invoke-model-titan")
     .WithOpenApi();
@@ -232,7 +236,8 @@ async Task PrepareAWSServer(IServiceProvider services)
 // Reroute the Bedrock API calls to our mock responses in BedrockTests. While other services use localstack to handle the requests,
 // we write our own responses with the necessary data to mimic the expected behavior of the Bedrock services.
 app.MapGet("guardrails/test-guardrail", (BedrockTests bedrock) => bedrock.GetGuardrailResponse());
-// For invoke model, we have one test case for each of the 6 suppported models.
+// For invoke model, we have one test case for each of the 7 suppported models.
+app.MapPost("model/us.amazon.nova-micro-v1:0/invoke", (BedrockTests bedrock) => bedrock.InvokeModelAmazonNovaResponse());
 app.MapPost("model/amazon.titan-text-express-v1/invoke", (BedrockTests bedrock) => bedrock.InvokeModelAmazonTitanResponse());
 app.MapPost("model/us.anthropic.claude-3-5-haiku-20241022-v1:0/invoke", (BedrockTests bedrock) => bedrock.InvokeModelAnthropicClaudeResponse());
 app.MapPost("model/meta.llama3-8b-instruct-v1:0/invoke", (BedrockTests bedrock) => bedrock.InvokeModelMetaLlamaResponse());

test/contract-tests/tests/test/amazon/awssdk/awssdk_test.py

@@ -550,6 +550,32 @@ def test_bedrock_get_guardrail(self):
             span_name="Bedrock.GetGuardrail",
         )
 
+    def test_bedrock_runtime_invoke_model_nova(self):
+        self.do_test_requests(
+            "bedrock/invokemodel/invoke-model-nova",
+            "GET",
+            200,
+            0,
+            0,
+            rpc_service="Bedrock Runtime",
+            remote_service="AWS::BedrockRuntime",
+            remote_operation="InvokeModel",
+            remote_resource_type="AWS::Bedrock::Model",
+            remote_resource_identifier="us.amazon.nova-micro-v1:0",
+            cloudformation_primary_identifier="us.amazon.nova-micro-v1:0",
+            request_response_specific_attributes={
+                _GEN_AI_SYSTEM: "aws.bedrock",
+                _GEN_AI_REQUEST_MODEL: "us.amazon.nova-micro-v1:0",
+                _GEN_AI_REQUEST_TEMPERATURE: 0.123,
+                _GEN_AI_REQUEST_TOP_P: 0.456,
+                _GEN_AI_REQUEST_MAX_TOKENS: 123,
+                _GEN_AI_USAGE_INPUT_TOKENS: 456,
+                _GEN_AI_USAGE_OUTPUT_TOKENS: 789,
+                _GEN_AI_RESPONSE_FINISH_REASONS: ["finish_reason"],
+            },
+            span_name="Bedrock Runtime.InvokeModel",
+        )
+
     def test_bedrock_runtime_invoke_model_titan(self):
         self.do_test_requests(
             "bedrock/invokemodel/invoke-model-titan",
@@ -982,6 +1008,7 @@ def _filter_bedrock_metrics(self, target_metrics: List[Metric]):
             "GET knowledgebases/test-knowledge-base",
             "GET knowledgebases/test-knowledge-base/datasources/test-data-source",
             "POST agents/test-agent/agentAliases/test-agent-alias/sessions/test-session/text",
+            "POST model/us.amazon.nova-micro-v1:0/invoke",
             "POST model/amazon.titan-text-express-v1/invoke",
             "POST model/us.anthropic.claude-3-5-haiku-20241022-v1:0/invoke",
             "POST model/meta.llama3-8b-instruct-v1:0/invoke",