REFRAG-style RAG (compress → sense/select → expand) — Single-file reference implementation，REFRAG框架通过压缩、选择和扩展流程显著提高RAG应用的效率，实现最高30.85倍的TTFT加速，并将LLM的有效上下文长度扩展16倍，解决了传统方法中上下文窗口过长的问题。
active 2025-09-10 → 2026-04-12 (UTC)
Activity over time
Daily event counts in the loaded window
Line chart, 215 days from 2025-09-10 to 2026-04-12. Pushes: 7 total, peak 2 in a day. Pull requests: 2 total, peak 1 in a day. Issues: 3 total, peak 1 in a day. Comments: 4 total, peak 3 in a day. Stars: 85 total, peak 13 in a day.
- Pushes
- Pull requests
- Issues
- Comments
- Stars
Stars, PRs, issues and forks are under-captured in the later part of this window. GH Archive progressively stopped capturing non-push events during 2026 — −95% or worse by the end of the window. Every series here except Pushes fades for that reason, so a decline above reflects the archive, not this repository. Pushes stay reliable throughout, so read them, and the contributor counts derived from them, as the real signal. Data health has the measurements.
Top contributors
Pushes, PRs, issues, reviews and comments — stars and forks excluded, so this is contribution rather than popularity
| Contributor | Contributions | Pushes | PRs | Comments |
|---|---|---|---|---|
| simulanics | 11 | 7 | 0 | 3 |
| missflash | 3 | 0 | 2 | 1 |
| simplzyu | 1 | 0 | 0 | 0 |
| Sumegh-git | 1 | 0 | 0 | 0 |
Recent activity
Latest issues, pull requests and releases
- Issue#4Sumegh-git2026-01-10 09:29[Feature Request] Add support for LoRA finetuning
- Issue comment#2simulanics2025-12-19 14:55Loss always 0 during cpt_next stage due to short tokens in data/cpt_train.jsonl
- Issue#2simulanics2025-12-19 14:45Loss always 0 during cpt_next stage due to short tokens in data/cpt_train.jsonl
- Pull request#3missflash2025-12-19 03:05
- Issue comment#3simulanics2025-12-19 03:05Add top-p (nucleus) sampling support to step-by-step decoder generation
- Issue comment#3missflash2025-12-19 01:59Add top-p (nucleus) sampling support to step-by-step decoder generation
- Pull request#3missflash2025-12-16 08:30
- Issue comment#1simulanics2025-09-28 02:11Reconstruction task: Shape mismatch when using decoder with inputs_embeds and labels
- Issue#1simplzyu2025-09-17 08:11Reconstruction task: Shape mismatch when using decoder with inputs_embeds and labels
Totals cover only the window loaded into ClickHouse and count events, not GitHub's lifetime totals — 85 stars here means stars gained during the window, not the repo's star count.