- Who it helps
- Developers evaluating retrieval-augmented generation, where answers use retrieved documents
- Availability
- Public MIT demo repository. Model and GPU requirements vary by variant.
Keep evidence verbatim
The encoder considers the question and passages together, then selects fragments without rewriting them. The author tests questions that require combining several facts; removing one connecting fact can still break an answer.
Treat compression scores as a starting point
The reported evaluation uses 400 MuSiQue questions and author-selected models. Training and comparison settings differ across compressors. Test retained facts and answer quality together on your own documents before adopting the method.
What you can do
- Inspect which supporting facts survive compression on one of your own retrieval examples. Open reference ↗
Where the evidence stops
README and maker discussion read November 14. No model run or independent reproduction verified.
Sources read. Public text checked November 14, 2026 UTC. No installation, model inference or full workflow executed.
Sources & further reading
GitHub · implementation
Chopper retrieval compressor ↗
README inspected. Verbatim extraction and MuSiQue comparisons are author evidence; no compression or answer generation run.
wiiiktor · page checked 2026-11-14
1 stars · observed 2026-11-14 19:13 UTC
More sources reviewed (2)
GitHub · discussion
Chopper maker discussion ↗
November 14 post, zero replies and one participant. Same maker as repository; not independent corroboration.
wiiiktor · page checked 2026-11-14
Hacker News · discussion
Maker launch discussion ↗
Original maker thread read; relative timestamps are not converted to exact launch dates.
Hacker News · page checked 2026-11-14
Coverage note: README and maker discussion read November 14. No model run or independent reproduction verified.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
Chopper keeps selected original text from retrieved passages to reduce the context an answering model reads.