The author has released Trimwise, an open-source Python library designed to fit source text into a specific context budget by removing irrelevant material while preserving evidence relevant to the current question.
- The library supports single-source trimming via `atrim` and multi-source trimming via `atrim_context`, which allows sources to compete for a shared budget.
- It selects existing source spans rather than generating summaries, providing provenance information to trace retained text back to its inputs.
- The default lexical path requires no model weights or LLM inference, though optional embedding-based modes are available for semantic and hybrid selection.
- Published benchmarks cover specific scenarios, with the author seeking feedback on edge cases such as lost qualifiers, negation, or revised facts.
The tool aims to address the trade-off where simple truncation discards needed context, offering a way to retain relevant evidence without generating new text.