Process a large batch of Groq inference requests asynchronously at reduced cost with the Batch API

Documented steps

1. Use the same key and auth header as synchronous calls: 'Authorization: Bearer <GROQ_API_KEY>'.
2. Build a JSONL file in which each line is one request object - chat completion, audio transcription, or audio translation. The file must not exceed 50,000 lines or 200MB, whichever limit binds first.
3. Upload it: POST https://api.groq.com/openai/v1/files as multipart/form-data with the file and a purpose field appropriate for batch. The response returns a file id.
4. Create the batch: POST https://api.groq.com/openai/v1/batches referencing the uploaded file id, the target endpoint (for example /v1/chat/completions), and a completion_window between 24h and 7d.
5. Poll GET https://api.groq.com/openai/v1/batches/{batch_id}. Status moves through validating, in_progress, finalizing and completed, or ends in failed or expired.
6. On completion the batch object exposes an output_file_id, and an error_file_id when some lines failed.
7. Download results: GET https://api.groq.com/openai/v1/files/{output_file_id}/content, which returns JSONL.
8. Correlate each result line back to its request using the per-line correlation id you set on submission (the OpenAI batch convention is custom_id - confirm the field name against the current batch docs before coding against it).
9. Fetch the error file too, and resubmit only the failed or unprocessed lines rather than the whole batch.
10. Official documentation: https://console.groq.com/docs/batch | https://console.groq.com/docs/api-reference