AWS Bedrock (Java)

Prerequisites

Complete the Java SDK setup first.

Installation

<dependency>
    <groupId>com.github.future-agi.traceAI</groupId>
    <artifactId>traceai-java-bedrock</artifactId>
    <version>main-SNAPSHOT</version>
</dependency>

implementation 'com.github.future-agi.traceAI:traceai-java-bedrock:main-SNAPSHOT'

You also need the AWS Bedrock Runtime SDK:

<dependency>
    <groupId>software.amazon.awssdk</groupId>
    <artifactId>bedrockruntime</artifactId>
    <version>2.25.0</version>
</dependency>

implementation 'software.amazon.awssdk:bedrockruntime:2.25.0'

Wrap the client

import ai.traceai.TraceAI;
import ai.traceai.bedrock.TracedBedrockRuntimeClient;
import software.amazon.awssdk.services.bedrockruntime.BedrockRuntimeClient;

TraceAI.initFromEnvironment();

BedrockRuntimeClient client = BedrockRuntimeClient.create();
TracedBedrockRuntimeClient traced = new TracedBedrockRuntimeClient(client);

InvokeModel (raw JSON)

The invokeModel API takes a raw JSON body. The wrapper parses the JSON to extract inputs and outputs based on the provider format.

import software.amazon.awssdk.core.SdkBytes;
import software.amazon.awssdk.services.bedrockruntime.model.*;

// Claude Messages format
String requestBody = """
    {
        "anthropic_version": "bedrock-2023-05-31",
        "messages": [{"role": "user", "content": "What is the capital of France?"}],
        "max_tokens": 1024
    }
    """;

InvokeModelResponse response = traced.invokeModel(InvokeModelRequest.builder()
    .modelId("anthropic.claude-3-haiku-20240307-v1:0")
    .body(SdkBytes.fromUtf8String(requestBody))
    .build());

String responseJson = response.body().asUtf8String();
System.out.println(responseJson);

Span created: “Bedrock Invoke Model” with kind LLM

The wrapper detects the provider from the model ID prefix and parses the JSON format accordingly:

Converse (typed API)

The converse API uses typed request/response objects instead of raw JSON. This is the recommended API for new integrations.

import software.amazon.awssdk.services.bedrockruntime.model.*;
import java.util.List;

ConverseResponse response = traced.converse(ConverseRequest.builder()
    .modelId("anthropic.claude-3-haiku-20240307-v1:0")
    .messages(List.of(
        Message.builder()
            .role(ConversationRole.USER)
            .content(List.of(ContentBlock.fromText("What is the capital of France?")))
            .build()
    ))
    .inferenceConfig(InferenceConfiguration.builder()
        .maxTokens(1024)
        .temperature(0.7f)
        .topP(0.9f)
        .build())
    .build());

String text = response.output().message().content().get(0).text();
System.out.println(text);

Span created: “Bedrock Converse” with kind LLM

What gets captured

Both APIs capture the same core attributes:

For invokeModel, the raw JSON body is stored in fi.raw_input and fi.raw_output. The wrapper does its best to extract structured messages from provider-specific JSON, but the raw JSON is always available as a fallback.

Accessing the original client

BedrockRuntimeClient original = traced.unwrap();