vercel/ai · error · InvalidArgumentError

Google batch input files must not exceed 2 GB.

Error message

Google batch input files must not exceed 2 GB.

What it means

Batch jobs upload requests as a single JSONL file to Google's Batch API, which caps input file size at 2 GB (googleBatchInputFileMaxBytes). experimental_doStartBatch checks the assembled Blob size before uploading and throws InvalidArgumentError if the file exceeds the limit, because Google would reject the upload anyway.

Source

Thrown at packages/google/src/google-batch.ts:266

    if (fileParts == null) {
      const { value } = await postJsonToApi({
        url: createUrl,
        headers,
        body: inlineBatchBody,
        failedResponseHandler: googleFailedResponseHandler,
        successfulResponseHandler: createJsonResponseHandler(
          googleBatchOperationSchema,
        ),
        abortSignal: options.abortSignal,
        fetch: this.batchConfig.fetch,
      });
      operation = value;
    } else {
      const inputFile = new Blob(fileParts, { type: 'application/jsonl' });
      // Blob snapshots the strings, so release the potentially large input array.
      fileParts.length = 0;
      if (inputFile.size > googleBatchInputFileMaxBytes) {
        throw new InvalidArgumentError({
          argument: 'requests',
          message: 'Google batch input files must not exceed 2 GB.',
        });
      }

      const { value: uploadUrl } = await postJsonToApi({
        url: `${this.getBaseOrigin()}/upload/v1beta/files`,
        headers: combineHeaders(headers, {
          'X-Goog-Upload-Protocol': 'resumable',
          'X-Goog-Upload-Command': 'start',
          'X-Goog-Upload-Header-Content-Length': String(inputFile.size),
          'X-Goog-Upload-Header-Content-Type': 'application/jsonl',
        }),
        body: {
          file: {
            display_name: `${displayName}-input`,
          },
        },

View on GitHub (pinned to 69428b1f8b)

Solutions

1. Split the requests into multiple batches, each under 2 GB, and start several batch jobs.
2. Remove or shrink inline data (use Cloud Storage URIs / file references instead of base64) to reduce JSONL size.
3. Trim prompts or move large static content out of per-request payloads (e.g. into context caching where supported).

Example fix

// before
await model.experimental_doStartBatch({ requests: allRequests }); // > 2GB
// after
for (const chunk of chunks(allRequests, 10_000)) {
  await model.experimental_doStartBatch({ requests: chunk });
}

const MAX_BYTES = 2 * 1024 ** 3;
function validateBatchSize(requests) {
  const size = new Blob(requests.map(r => JSON.stringify(r) + '\n')).size;
  if (size > MAX_BYTES) throw new Error(`Batch payload ${size} bytes exceeds 2 GB; split into multiple batches.`);
  return size;
}

null

try {
  await model.experimental_doStartBatch({ requests });
} catch (e) {
  if (e?.name === 'AI_InvalidArgumentError' && /2 GB/.test(e.message ?? '')) {
    // split requests into chunks and start multiple batch jobs
  } else throw e;
}

When it happens

Trigger: Calling `model.experimental_doStartBatch({ requests })` (or the batch helper) with a request array whose serialized JSONL exceeds 2,147,483,648 bytes - typically very large prompts, inline file/image data, or simply too many requests.

Common situations: Batch-embedding large document corpora with inline attachments; long-running backfills of thousands of requests with big system prompts; accidentally embedding base64 media in batch requests.

Related errors

• targetLanguage is required for translation model '${this.mod
• The Gemini Live translation API only supports 16kHz 16-bit P
• maxEmbeddingsPerCall must be greater than 0
• maxInputBytesPerCall must be greater than 0
• No image generated.

AI-assisted analysis of vercel/ai@69428b1f8b (2026-08-30). Data as JSON: /api/errors/a55134dc9c94aec8. Report an issue: GitHub.