Limits & Timeouts
Two numeric fields decide whether a batch comes back whole. They are the most common cause of translations that quietly look like the original.
LLM Max Token
aireviewtranslator/general/llm_max_token
Digits only. Caps the model's output for a single call. Leave it empty to use the provider's own default.
Size it against your batch
The bulk command sends 20 reviews per call by default and expects a JSON array covering all 20 back. Too small a ceiling truncates that JSON, the reply fails to parse, and every review in the batch falls back to its original text — with no error.
| Batch size | Rough output need |
|---|---|
| 1–5 reviews | A few hundred tokens |
| 20 (default) | 1,500 is a reasonable start |
| 50+ | 3,000+, or leave it empty |
Raise the batch, raise the ceiling
--batch and LLM Max Token move together. Changing one without the other is what produces the silent fallback described in Fallback Behaviour.
Ollama Request Timeout
aireviewtranslator/general/llm_timeout
Ollama only. Minimum 30, default 300 seconds.
The request is non-streaming: Ollama returns nothing until the entire batch has been generated. The timeout therefore has to cover the whole batch, not one review.
Sizing guidance is on Timeouts & Batch Size.
Neither field affects the queue path
The observer path translates one review at a time, so the default ceiling is almost always enough there. These settings matter most for bulk runs.