Document input not supported for token counting

[azavisha-snap]

Describe the bug

There is very little information about the new token counting feature for Bedrock. It seems to work for basic text, which is cute, but the difficult part is documents, and it doesn't work for any kind of documents. I know that Claude 4 supports token counting, at least through their API, but Bedrock is throwing an exception.

Regression Issue

• Select this option if this issue appears to be a regression.

Expected Behavior

I expect token counting to work for Claude 4, including documents such as txt, md, doc, docx, csv, xls, xlsx, pdf.

Current Behavior

Exception for any document: "Document input not supported for token counting".

Reproduction Steps

bedrockRuntimeClient.countTokens({
	it.modelId('anthropic.claude-sonnet-4-20250514-v1:0')
	it.input({
		it.converse {
			it.messages([
				Message.builder()
					.content([
						ContentBlock.fromDocument({
							it.name('document')
							it.source {
								it.text('text')
							}
						})
					])
					.role(ConversationRole.USER)
					.build()
			])
		}
	})
}).inputTokens()

Possible Solution

No response

Additional Information/Context

No response

AWS Java SDK version used

2.32.29

JDK version used

11

Operating System and version

MacOS 15.0.1

[bhoradc]

Hello @azavisha-snap,

Thank you reporting the issue. I am getting the same exception from the service. Here's the minimal reproducible code sample I have used for Java SDK v2.x.

Main.java

import software.amazon.awssdk.regions.Region;
import software.amazon.awssdk.services.bedrockruntime.BedrockRuntimeClient;
import software.amazon.awssdk.services.bedrockruntime.model.*;
import java.nio.file.Files;
import java.nio.file.Paths;

public class Main {
    
    public static void main(String[] args) {
        String documentPath = "sample.txt";
        int tokenCount = countDocumentTokens(documentPath);
        System.out.println("Document: " + documentPath);
        System.out.println("Token count: " + tokenCount);
    }
    
    public static int countDocumentTokens(String documentPath) {
        try (BedrockRuntimeClient client = BedrockRuntimeClient.builder()
                .region(Region.US_EAST_1)
                .build()) {
            
            byte[] documentBytes = Files.readAllBytes(Paths.get(documentPath));
            
            CountTokensResponse response = client.countTokens(CountTokensRequest.builder()
                    .modelId("anthropic.claude-sonnet-4-20250514-v1:0")
                    .input(CountTokensInput.builder()
                            .converse(ConverseTokensRequest.builder()
                                    .messages(Message.builder()
                                            .role(ConversationRole.USER)
                                            .content(
                                                ContentBlock.fromText("Count tokens for this document"),
                                                ContentBlock.fromDocument(
                                                    DocumentBlock.builder()
                                                        .name("document")
                                                        .format(DocumentFormat.TXT)
                                                        .source(DocumentSource.fromBytes(
                                                            software.amazon.awssdk.core.SdkBytes.fromByteArray(documentBytes)))
                                                        .build()))
                                            .build())
                                    .build())
                            .build())
                    .build());

            return response.inputTokens();
        } catch (Exception e) {
            System.err.println("Error counting document tokens: " + e.getMessage());
            return -1;
        }
    }
}

I have reached out to the internal bedrock service team to know whether document input support for CountTokens API is a pending feature. Will post here once I hear back from the team.

Regards,
Chaitanya