The results of a finished batch, read from its provider. OpenAI keeps them 30 days from when it wrote them, at the end of the batch (jobs module docstring), and Anthropic 29 days after the batch was created; Mistral states that batch results do not expire. Amazon Bedrock and Vertex AI results are read from Eden AI's own storage. Once the provider no longer has them the answer is 410. An Anthropic, Bedrock or Vertex AI request's result is translated to a chat completion.

The response is streamed. The download is not: llm__abatch_results reads both files whole, so a request holds them in memory, bounded by the create's line cap (LLM_BATCH_MAX_LINES). The lines are validated once there and written out as dicts.