Localize one Sume Format into six languages: one bulk item each

Send one bulk item per language to the same Format, with the language in input, and let the queue run them under one concurrency window. Six languages is six items and six independent runs, each with its own receipt, spend cap and media. Whether a Format reads a language key is up to its recipe; the docs' live-commerce example reads vo_language.

The queue shape

A bulk request takes concurrency (1 to 16) and items (1 to 100). Every item is an ordinary run body. The Format reads the keys of input that it knows, and passes the others along as context, so check your Format's recipe for the exact key before you rely on it.

{
  "concurrency": 3,
  "items": [
    { "input": { "sku": "8823", "vo_language": "es" } },
    { "input": { "sku": "8823", "vo_language": "de" } },
    { "input": { "sku": "8823", "vo_language": "fr" } }
  ]
}

Keys, schemas and checks

Give each language its own key so a retry of one language cannot be mistaken for another. Bind the same output_schema on every item, because the schema is per run and not inherited.

Checking each language

Read each child receipt, not only the queue. The queue says an item is completed, while the receipt shows output, artifacts[] and output_error. A completed run with a null output is a different case from a clean one, and a mixed batch of six languages can contain both.

Listen to the voice track of every language before publishing. The Format may read your language key, but only a speaker can judge pronunciation, brand terms and line length, which often grows in languages such as German.

Tradeoffs

Sume does not translate for you in the queue; each run is a full production, so localization costs six runs, not one run plus six cheap edits. If only the voice and captions change, a continuation with previous_run_id on the finished run may cost less, because it reuses what the run already made. Have a native speaker listen to a pilot of one language before you queue the rest.