Translation engine
Segment size, concurrency, temperature, retries and prompt caching.
Segment size
CHUNK_TOKEN_SIZE defaults to 1000 and accepts values from 200 to 10000. A lower value makes recovery easier and reduces context; a higher value can improve continuity but increases cost, latency and overflow risk.
Concurrency
CONCURRENCY defaults to 1 and accepts values from 1 to 32. Increase it gradually: the provider, local server, VRAM and rate limit must all support the parallel load.
Temperature
The default is 0.15, within a range of 0 to 2. For faithful and reproducible translation, generally keep it low. A high temperature increases variation and can harm terminology consistency.
Retries and timeout
MAX_RETRIES=3: maximum number of controlled attempts.REQUEST_TIMEOUT=180.0: timeout in seconds, from10to1800.
A disconnection or provider outage may pause the project so that results already written remain safe.
Prompt caching
A project can enable a prompt-cache flag. Its effectiveness depends on the provider and model; it does not guarantee savings when the selected API does not support the mechanism.