Batch API result file download fails silently for large batches (>~200MB output)

API Bugs
, ,
1
  1. Symptom: Submitted 50,000 embedding requests using the Batch API, resulting in an output file of approximately 1GB. The download fails with the following error: peer closed connection without sending complete message body (received X bytes, expected Y bytes).
  2. Conditions to Reproduce: The number of batch lines approaches the 50,000 limit, and the output file size exceeds ~200-300MB.
  3. Documentation Gap: The official documentation does not mention any impact of output file size on download stability.
  4. Workaround: Limit each batch to under 10,000 lines (keeping the output around ~200MB) to ensure stable downloads.
  5. Environment: openai-python SDK, macOS. The issue is reproducible using both with_streaming_response.content() and the standard requests library.

Not sure if only me met this issue…which i didnt see in the developer docs.

thank you in advanced.

1 Like
2

Hi and welcome to the community!

Have you tried to download the batch via the platform? This should get you unstuck for now and you can retrieve your embeddings.

https://platform.openai.com/batches

1 Like