ApartureBringing the arXiv into focus
Multi-stage research paper discovery and analysis using LLMs
Multi-stage research paper discovery and analysis using LLMs
Aparture is a local web app for daily arXiv monitoring using LLMs. You write a profile describing your research interests in plain English, pick a set of arXiv categories to watch, and each run pulls down new preprints, scores them, reads the top PDFs, and produces a briefing — a short editorial summary, a handful of thematic groupings, and per-paper notes tuned to your profile.
It's designed to be used over weeks rather than once: you star papers you liked, dismiss ones that missed the mark, and comment where you have specific reactions. Those signals feed into a profile-refinement flow that proposes edits you can accept or reject individually, so the briefings gradually converge on what you actually want to see.
Write your interests as plain prose, not a keyword list. No taxonomies, no schemas, no tuning sliders — the LLM reads what you wrote and uses that to score and filter.
Each run produces one structured reading piece: an editorial lead, themes, and per-paper cards. Designed to help you decide what to actually open, not just produce a list.
Star, dismiss, or comment on what you see. Those signals accumulate, and the suggest-improvements flow proposes concrete profile edits you can accept or reject per-change.
Works with Anthropic Claude, OpenAI, and Google Gemini. Mix providers across stages — cheap model for filtering, stronger model for synthesis. All keys stay in your local .env.local.
Node, the repo, optional Playwright. About 5 minutes.
STEP 02Pick a provider, create a key, paste into .env.local.
Dry run (free) and Minimal API Test (~$0.20–$1 on paid tiers).
A narrated 10-minute walk through your first real briefing, end to end.
Already have Aparture running? The Guide covers daily use — running briefings, reading them, giving feedback, writing a profile, refining over time — Under the Hood goes deeper into how the pipeline and briefings actually work, and Reference is the lookup surface for env vars, prompt files, and troubleshooting symptoms.