pub(crate) const LOW_WATERMARK_PCT: usize = 75;Expand description
Low watermark the trim aims for, as a percentage of the request budget.
Once truncation is triggered (payload over limit_chars), the stage trims
past the budget down toward this fraction of it, buying several turns of
headroom in which follow-up requests need no new elisions — the property
that keeps the forwarded prompt prefix stable for llama.cpp’s KV cache.
See the module docs for why the savings target is also quantized.