Split Text (wtpsplit)

How it fits

Typical backends

Caveats

• I/O contractOutput is one entry per sentence and one list per input, always; empty or whitespace-only input yields the empty list, never a dropped message. Group the results downstream when paragraph-sized chunks are needed, for example to fill a retrieval context window or a prompt budget.
• AccuracyThe model is language-agnostic across major scripts, but heavily mixed-language text or highly unusual punctuation can degrade split quality without an error. Raise threshold for fewer, longer sentences and lower it for more, shorter ones.
• CostThe SaT model is small (~0.2B) and runs in full precision; the default device is ``cpu``, which splits a typical message in tens of milliseconds, so no GPU is required. The SaT-3L-SM weights and the XLM-RoBERTa subword tokenizer they wrap are staged offline from platform files and loaded locally, so there is no runtime download and no HuggingFace token.
• State lifetimedevice is captured once at startup; changing it requires a redeploy. threshold and strip_whitespace are re-read every message.