Chopper: with and without
A question that needs two facts from 20 retrieved Wikipedia paragraphs. Without Chopper, Qwen3-14B reads all 20 paragraphs. With Chopper, it reads only the exact quotes Chopper cut out, about 12× fewer tokens. Same model, same prompt.
GitHub · model · po polsku · pip install chopper-rag · MIT, commercial use allowed
20 example questions: Qwen3-14B's answers
Green = correct, red = wrong. Sorted: questions Chopper got right first. Click a row to see it below, with the paragraphs and chops.
| Question | Without Chopper | With Chopper |
|---|
- Chopper cuts the chops in your browser (the model, ~450 MB, starts loading as soon as the page opens; cutting takes 10–60 s).
- Qwen3-14B then answers live, twice: once reading all 20 paragraphs, once reading only the chops.
Retrieved paragraphs (20): what Qwen reads without Chopper
Without Chopper
With Chopper
Data: MuSiQue 2-hop development questions (CC BY 4.0), each with the 20 paragraphs MuSiQue provides. Answers by Qwen3-14B (no thinking, answer only, temperature 0), judged against the gold answer and its aliases by an LLM judge. The 20 examples were drawn at random, stratified to match the result on all 100 two-hop questions of our hop study; chops are the released Chopper's raw output (what Qwen read), so some start or end mid-sentence. For your own questions Chopper runs here in your browser (base model, whole sentences) and Qwen3-14B answers through a small proxy; nothing is stored. Full results. * Time: an estimate for one RTX 3090 running Qwen3-14B (4-bit): about 1,800 prompt tokens per second plus about 0.25 s to write a short answer; with Chopper add 0.16 s for Chopper itself (measured 155 ms per lookup). Live answers come through OpenRouter, so their real time depends on the network.