arXiv drops hundreds of papers a night. Say what you care about, and this page ranks the whole night against it.
- 1 Haul the stones Fetch tonight's papers.
- 2 Filter Say what interests you.
- 3 Assay Keep the pay dirt.
Runs in this tab. Text you score is embedded, not stored. Small print.
What is sent. Your touchstones, core samples and weights stay in this tab and survive a reload. The one thing that leaves is text you want scored: the pick runs on a server, so that text is sent there to be embedded, and is not stored.
The shared cache. Published text (arXiv abstracts, core-sample abstracts, preset touchstones) also goes into a shared vector cache, keyed by a hash, so nobody pays to cut the same phrase twice. A touchstone you type yourself never enters that cache, because reads on it are public and a hash of a short phrase is one dictionary away from the phrase.
Your own pick or relay. Set window.ARXAVE_EMBED to any OpenAI-shaped /v1/embeddings (LM Studio, Ollama, a paid endpoint) and text is embedded there instead, with the shared cache skipped. Set window.ARXAVE_RELAY the same way for the arXiv fetch. Or fork it (source).
Touchstone length. One word and a paragraph are not directly comparable, since longer text is more specific. The per-row weights are the mitigation.
What is real today. Scouting arXiv, embedding abstracts, and ranking on touchstone similarity with live re-blending. The rush (Scirate) sits behind a Cloudflare challenge and stays inactive.