Expand description
Neural context compression — trained models replacing heuristic filters.
Feature-gated under #[cfg(feature = "neural")].
When an ONNX model is present, switches from heuristic to neural scoring.
Falls back gracefully to heuristic mode when no model is available.
Modules§
- attention_
learned - Learned attention weights replacing the heuristic U-curve.
- cache_
alignment - KV-Cache alignment for commercial LLM prompt caching.
- context_
reorder - line_
scorer - Neural line importance scorer using ONNX inference via rten.
- token_
optimizer - Token-optimal encoding based on empirical lab results.