Class AnthropicBedrockSDKClient

java.lang.Object

ai.redouble.nucleo.harness.llm.AbstractRateLimitedClient

ai.redouble.nucleo.harness.llm.AbstractLLMClient<com.anthropic.models.messages.ContentBlockParam>

ai.redouble.nucleo.providers.anthropic.AnthropicSDKClient

ai.redouble.nucleo.providers.bedrock.anthropic.AnthropicBedrockSDKClient

All Implemented Interfaces:

Client, LLMClient

Direct Known Subclasses:

AnthropicBedrockMantleSDKClient

public class AnthropicBedrockSDKClient
extends AnthropicSDKClient

Anthropic Claude API client using AWS Bedrock as the backend. Uses the official Anthropic Java SDK with BedrockBackend.

Key differences from direct API (AnthropicSDKClient):

• Uses AWS credentials instead of Anthropic API key
• Model IDs use Bedrock format: anthropic.claude-opus-4-5-20251101-v1:0
• Rate limiting uses AWS throttling (ThrottlingException)

Credentials and region per BedrockClients: AWS's own resolution (the standard variables, a profile, or the role the process runs under), or the deployment's store records under BedrockClients.SECRET_ID and BedrockClients.REGION_ID.

Since:

0.1 (2025-12-02)

Author:

Andrey Santrosyan

See Also:

• AnthropicSDKClient
• AnthropicModelResolver.resolve(ModelSpec)

Nested Class Summary

Nested classes/interfaces inherited from class AbstractRateLimitedClient

AbstractRateLimitedClient.RateLimitedOperation<T>

Field Summary

Fields inherited from class AnthropicSDKClient

anthropicModel, client, SECRET_ID, SHAPE

Fields inherited from class AbstractLLMClient

formatter, temperature

Fields inherited from class AbstractRateLimitedClient

model

Constructor Summary

Constructors

Modifier	Constructor	Description
	AnthropicBedrockSDKClient()
protected	AnthropicBedrockSDKClient(com.anthropic.backends.Backend backend)	Shared by the Bedrock endpoint variants.

Method Summary

Modifier and Type	Method	Description
protected static software.amazon.awssdk.auth.credentials.AwsCredentialsProvider	awsCredentialsProvider()	AWS credentials for either Bedrock surface, resolved exactly as the runtime client resolves them (BedrockClients.credentialsProvider()).
protected static software.amazon.awssdk.regions.Region	bedrockRegion()	The deployment's Bedrock region, the same one for either surface.
protected <T> LLMResponse<T>	doSingleResponse(LLMRequest<T> request, PreparedConversation prepared)	The Anthropic call, with a Bedrock reading of one failure the shared client cannot give: an unresolvable endpoint hostname means the model is not served in this deployment's region, because the region built the hostname (bedrock-mantle.<region>.api.aws).
protected <T> void	doStreamResponse(LLMRequest<T> request, LLMResponse<T> response, PreparedConversation prepared, Consumer<StreamChunk> chunkHandler)	Streaming implementation.
APIDialect	getDialect()	Gets the API dialect for this client.
protected boolean	is429Error(Exception e)	Bedrock surfaces two different 429 shapes depending on which layer hits the limit.
void	setModel(ModelSpec modelInfo)	Uses Bedrock model ID format.

Methods inherited from class AnthropicSDKClient

accountIdentifier, attachThinkingConfig, buildEncoders, buildMessageCreateParams, close, countTokensExact, createFormatter, extractRateLimitInfo, getAnthropicModel, isCachingSupported, isOverloadError, isQuotaError, isServerError, setAnthropicModel, textWrapper

Methods inherited from class AbstractLLMClient

base, createOutgoingMessage, encodeBlock, estimateTokens, findUnknownHost, getCallerInfo, getFormatter, getModelIdentifier, getTemperature, isPromptTooLongError, prepareConversation, reasoningEffortFor, resolveWireMaxTokens, setTemperature, singleResponse, streamResponse, updateRateLimitsFromHeaders, validateEstimatedTokens

Methods inherited from class AbstractRateLimitedClient

executeWithRateLimit, getModel, messageChainContains, quotaExhausted

Methods inherited from class Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Methods inherited from interface Client

getModel

Constructor Details

AnthropicBedrockSDKClient

public AnthropicBedrockSDKClient()

AnthropicBedrockSDKClient

protected AnthropicBedrockSDKClient(com.anthropic.backends.Backend backend)

Shared by the Bedrock endpoint variants. AnthropicBedrockMantleSDKClient passes a BedrockMantleBackend here; everything downstream of the backend - throttle classification, model resolution, rate-limit header extraction - is identical between the two surfaces, so it lives once, in this class.

Method Details

awsCredentialsProvider

protected static software.amazon.awssdk.auth.credentials.AwsCredentialsProvider awsCredentialsProvider()

AWS credentials for either Bedrock surface, resolved exactly as the runtime client resolves them (BedrockClients.credentialsProvider()). An unreadable secret surfaces as the store's own unchecked failure: this client cannot be built at all.

bedrockRegion

protected static software.amazon.awssdk.regions.Region bedrockRegion()

The deployment's Bedrock region, the same one for either surface.

doSingleResponse

protected <T> LLMResponse<T> doSingleResponse(LLMRequest<T> request,
 PreparedConversation prepared)

The Anthropic call, with a Bedrock reading of one failure the shared client cannot give: an unresolvable endpoint hostname means the model is not served in this deployment's region, because the region built the hostname (bedrock-mantle.<region>.api.aws).

Overrides:

doSingleResponse in class AnthropicSDKClient

Parameters:

request - the LLM request

prepared - the prepared conversation, ready for provider-specific encoding

Returns:

the LLM response

doStreamResponse

protected <T> void doStreamResponse(LLMRequest<T> request,
 LLMResponse<T> response,
 PreparedConversation prepared,
 Consumer<StreamChunk> chunkHandler)

Description copied from class: AnthropicSDKClient

Streaming implementation. Parses message_start for initial input + cache counters and the message ID; message_delta for the final output_tokens and stop_reason; content-block deltas for text. Populates the response via LLMResponse.setAdditiveUsage(int, Integer, Integer, int) once the stream completes. rateLimitInfo is not populated on streaming because the SSE transport doesn't expose HTTP response headers.

Overrides:

doStreamResponse in class AnthropicSDKClient

getDialect

public APIDialect getDialect()

Description copied from class: AbstractLLMClient

Gets the API dialect for this client. Used to determine encoding strategies and compatibility.

Specified by:

getDialect in interface LLMClient

Overrides:

getDialect in class AnthropicSDKClient

Returns:

the API dialect

setModel

public void setModel(ModelSpec modelInfo)

Uses Bedrock model ID format.

Specified by:

setModel in interface Client

Overrides:

setModel in class AnthropicSDKClient

is429Error

protected boolean is429Error(Exception e)

Bedrock surfaces two different 429 shapes depending on which layer hits the limit. Anthropic's own TPM/RPM check is proxied through Bedrock as com.anthropic.errors.RateLimitException with message "429: Too many tokens..." - this is the common case, and the parent's message-based detector already catches it. AWS-side throttling appears as ThrottlingException / ServiceQuotaExceededException in the cause chain, which we detect here. Failing to catch either of these makes the job terminate hard instead of riding through the dispatcher's transparent retry loop.

Overrides:

is429Error in class AnthropicSDKClient

Parameters:

e - the exception to check

Returns:

true if this is a rate limit error