Exact BatchInference job gcsUri output path is missing

[GolanLevy]

Similar to #696, I have no way to get the exact output file path of a batch job response file.

The gcsUri field contains the output URI provided in the request, which is not very helpful as it is only a prefix.
The actual file resides in a subfolder, named {gcsUri}/prediction-model-{timestamp}/predictions.jsonl.
I was not able to correlate the timestamp to any field. It is very similar to the createTime field but not exact.

For example, if I set the gcsUri to be gs://bucket_name/output/ the output file path will be something like gs://bucket_name/output/prediction-model-2025-12-26T20:29:02.829566Z/predictions.jsonl

However, the batch inference UI web console does have an Export Location field that holds the correct output file path. This hopefully means that you can easily expose that to the Java SDK as well.

Can you please fix that or provide any proper workaround?
Currently I have to list all files under the gcsUri prefix and hope there is going to be only one response file.

[hemasekhar-p]

Hi @GolanLevy , Thank you for bringing this issue to our attention. To assist us in effectively investigating and reproducing it, please provide a minimal, runnable code example, including the exact API call, Full request payload, which Model you are using and the complete error message?

[GolanLevy]

Hi @hemasekhar-p, here is the code, assuming job named "jobId" had already finished.
This is not related to any specific Model.

Client client = Client.builder()
                .project("projectId")
                .location("us-east4")
                .vertexAI(true)
                .build();

BatchJob batchJob = client.batches.get("jobId", GetBatchJobConfig.builder().build());

System.out.println(batchJob.dest().get().gcsUri().get());

As explained above, there is no error.
The output from the code above does not return the exact path to the response file.

[hemasekhar-p]

Hi @GolanLevy, thank you for the additional details. I have successfully reproduced the issue and escalated it to our engineering team for further investigation. Until the engineering team fixes this, you can try the below workaround and continue without being blocked.

package com.example.repros.jan2026;
import com.google.api.gax.paging.Page;
import com.google.cloud.storage.Blob;
import com.google.cloud.storage.Storage;
import com.google.cloud.storage.StorageOptions;
import java.util.Comparator;
import java.util.stream.StreamSupport;

public class BatchOutputWorkaroundRepro {

    public static void main(String[] args) {
        String bucketName = "BUCKET_NAME";
        String basePrefix = "";
       
        String finalPath = findLatestVertexOutputPath(bucketName, basePrefix);

        if (finalPath != null) {
            System.out.println("SUCCESS: Found actual output path: " + finalPath);
            System.out.println("We can use this path to download predictions.jsonl");
        } else {
            System.out.println("FAILURE: No prediction folder found. Check bucket name or job status.");
        }
    }

    public static String findLatestVertexOutputPath(String bucketName, String prefix) {
        Storage storage = StorageOptions.getDefaultInstance().getService();

        String searchPrefix = (prefix.isEmpty() || prefix.endsWith("/")) ? prefix : prefix + "/";

        Page<Blob> blobs = storage.list(bucketName,
                Storage.BlobListOption.prefix(searchPrefix),
                Storage.BlobListOption.currentDirectory());

        return StreamSupport.stream(blobs.iterateAll().spliterator(), false)
                .filter(Blob::isDirectory)
                .filter(b -> b.getName().contains("prediction-")) 
                .max(Comparator.comparingLong(b -> {                   
                    return b.getCreateTime() != null ? b.getCreateTime() : 0L;
                }))
            .map(b -> "gs://" + bucketName + "/" + b.getName())
            .orElse(null);
    }
}