`predict_entities` truncates silently -- expose truncation programmatically

[girishponkiya]

On gliner 0.2.27, input past config.max_len is truncated with only a UserWarning (data_processing/processor.py:513-515). In a service, that is invisible: warnings do not reach structured logs and are deduplicated per process, so the caller gets a normal-looking result with ~70% of a news article never read (gliner_medium_news-v2.1, max_len=296).

Ask:
A per-call signal, e.g. predict_entities(..., return_truncation_info=True) returning {"truncated": True, "num_tokens": N, "max_len": M}, or a truncation="error" mode (#231 asked for the latter, unanswered). Today I count model.data_processor.words_splitter(text) against model.config.max_len, but both are internals.

Doc ask:
Truncation applies to the text tokens before the entity-type prompt is prepended, so the label count never reduces the text budget -- could you confirm and document this? The natural assumption for a uni-encoder is the opposite, and the recurring limit confusion (#183, #275, #185; #244 relatedly) suggests an "input limits" doc section would help.

[girishponkiya]

For anyone landing here: there is a construction-time override -- GLiNER.from_pretrained(..., max_length=N) rewrites config.max_len before the processor is built, and it works (verified on gliner_medium_news-v2.1: with max_length=512, entities past word 296 are found with high scores).

But it is orthogonal to this issue's ask. It moves the ceiling; it does not tell the caller when the ceiling was hit. Whatever max_len is in effect, truncation is still signalled only by the per-process UserWarning. And raising it past the fine-tuning length trades silent truncation for out-of-distribution inference (#183's caveat: "performance can start to degrade"), so it is not a general fix -- a per-call truncated signal is still needed either way.

[Ingvarstep]

Hi @girishponkiya, thank you for pointing out to this, we made the required clarrifications in the docs and considering implementing your proposition.