A reading desk with an AI in the margin.
Citation-aware. Quiet until you ask.
Your keys. Your model. Local-first.
一张书桌,AI 住在页边。
识别引文。你不问,它不响。
你的密钥,你的模型,本地优先。
1.Reading papers without losing your place — or the thread of why you started.
2.A Chrome extension that turns PDF tabs into a quiet, AI-assisted reading desk.
1.读论文时不再丢失批注、位置,以及最初翻开它的理由。
2.一个把 PDF 标签变成安静的、AI 协助的阅读桌的 Chrome 扩展。
You open a preprint at 10:42 a.m. By 6:13 p.m. it's still unread, wedged between thirty tabs that all promised the same thing.
PaperFlow is a reading desk — the AI lives in the margin. You highlight; it remembers. You ask; it answers from the passages it read.
早上 10:42 打开 preprint,下午 6:13 还没读,挤在三十个同样许诺的标签之间。
PaperFlow 是一张书桌——AI 住在页边。你划线,它记住;你提问,它从你读过的段落里回答。
3.2 Routing. We introduce SMoDR, a sparse mixture of discrete retrievers that decouples retrieval from reasoning at the token level[3]. By routing only the top-k experts per query token, SMoDR achieves a 2.4× speedup over dense baselines while improving exact-match by 3.1 points[7].
The router is trained jointly with a contrastive objective, similar in spirit to prior work on sparse MoE[12], but with a quantized expert identity that compresses the dispatch tensor by an order of magnitude.
Touches only arXiv and PDF tabs.
只触碰 arXiv 与 PDF 标签。
arXiv, hosted PDF, or local file://.
arXiv、托管 PDF,或本地 file://。
Paste a key. Stored locally.
粘贴密钥,仅存本地。
Highlight, chat, search — one ⌘K away.
划线、对话、检索——只在 ⌘K 之外。
Local — chrome.storage.local.
本地 — chrome.storage.local。
Optional. End-to-end encrypted.
可选,端到端加密。
None. No analytics, no telemetry.
无。无分析,无遥测。
MIT. Self-host the sync server.
MIT 开源,可自建同步。
Free for personal reading. You pay your model provider directly.
个人阅读永久免费。模型费用直付提供方。
Any OpenAI-compatible endpoint: Claude, Gemini, DeepSeek, Ollama, LM Studio.
任何 OpenAI 兼容接口:Claude、Gemini、DeepSeek、Ollama、LM Studio。
No. Traffic goes browser → provider. Nothing through us.
看不到。流量从浏览器直达提供方。
Chrome & Edge now. Firefox soon. Safari later.
现支持 Chrome 与 Edge。Firefox 在做。