Batch Size
Batching is the difference between 18 API calls and 350. It is also the setting most likely to cause a silent failure, so it gets its own page.
The trade-off
--batch | Trade-off |
|---|---|
1–5 | Slowest and most calls, but the most robust. Good for debugging or a slow local model. |
20 (default) | Sensible for most hosted providers. |
50+ | Fewest calls. Needs a generous token ceiling or the JSON reply is truncated. |
The arithmetic
With 350 reviews in one store view:
| Batch | API calls | Reduction |
|---|---|---|
| 1 | 350 | — |
| 20 | 18 | 95% |
| 50 | 7 | 98% |
The command prints this reduction in its summary.
Batch size and max tokens move together
This is where translations silently fail
The model must return the whole batch as one JSON array. If LLM Max Token is too small for the batch, the reply is cut off, the JSON fails to parse, and every review in that batch falls back to its original text — with no error at all.
Raise the batch and you must raise the token ceiling, or clear it and use the provider default.
On Ollama, the timeout too
A large batch also needs a higher Ollama Request Timeout, because the request is non-streaming and returns nothing until the whole batch is generated. See Timeouts & Batch Size.
Finding your number
Start at the default, then test upward on one store view:
php bin/magento Translate:All -s 2 -b 40
Check the summary for warnings, then check for fallbacks:
SELECT COUNT(*)
FROM ai_review_translation_list t
JOIN review_detail d ON d.review_id = t.review_id
WHERE t.store_id = 2 AND t.review_message = d.detail;
A jump in that count when you raise the batch means the ceiling is too low. Back off, or raise LLM Max Token.
An invalid value is not fatal
A --batch of 0 or a negative number prints a warning and falls back to 20.