Paper Radar / AI 论文雷达

A static AI paper discovery site focused on scanability, bilingual reading, and lightweight browsing.

Paper Radar tracks recent AI papers, groups them into clearer discovery surfaces, and publishes everything as a pure static site. The product is built around a simple principle: default to light, scannable information first; let deeper reading and heavier search happen on demand.

This is not a general academic database UI and not a backend-heavy paper platform.

It is a curated static discovery layer for recent AI research:

• Home gives a quick editorial overview of what is new and what is worth checking first.
• Latest is for date-based browsing of newly collected papers.
• Rising Signals highlights papers gaining attention based on multiple signals, not just raw recency.
• Search starts with lightweight metadata search and can optionally unlock heavier full-text search.
• Paper Detail provides bilingual summaries, metadata, and source links.

The site is fully static and deploys to GitHub Pages.

The current product direction is:

• lightweight by default
• strong scanability on mobile and desktop
• bilingual /en and /zh routes
• human-readable grouped categories instead of raw taxonomy everywhere
• ranked surfaces that explain why something is worth opening
• optional depth instead of forcing every user to download or parse everything up front

The data pipeline is build-time and repository-driven.

1. scripts/fetch.js pulls and merges recent papers.
2. scripts/summarize.py generates bilingual summaries and supporting metadata.
3. scripts/pull_citation_signals.py stores daily citation snapshots.
4. scripts/pull_hf_signals.py stores Hugging Face Daily Papers snapshots.
5. scripts/compute_heat_scores.py computes ranking signals for the recent window.
6. scripts/export_heat_json.py exports frontend-facing heat metadata.
7. GitHub Actions commits refreshed src/data/**, builds the site, and deploys dist/.

Core content lives in:

• src/data/papers.json
• src/data/paper_heat_scores.json
• src/data/field_heat_topn.json
• src/data/heat_score_meta.json
• src/data/signals/**

Rising Signals and other ranked surfaces use a multi-signal heat model over the recent 180-day window.

Base score combines:

• field-normalized citation signal
• Hugging Face community buzz
• venue authority
• freshness decay by age bucket
• a reserved code signal slot

Burst bonus adds extra lift when enough history exists, especially for citation deltas and recent community appearance.

The system also tracks warmup state:

• cold: no citation history yet
• warmup: 1-6 daily snapshots
• ready: 7+ daily snapshots, citation burst is fully enabled

That state is exported in heat_score_meta.json so the site can distinguish between early and mature ranking periods.

Search is intentionally split into two levels:

• default quick search over lighter metadata fields
• optional full-text search, activated by the user when they want deeper recall

This keeps the default experience lighter for a static site while still allowing deeper search when needed.

• Astro 4
• React 18 islands
• Tailwind CSS 3
• TypeScript
• Framer Motion

The frontend is static-first. Interactive React islands are used only where they materially improve browsing or search.

The project deploys through GitHub Actions to GitHub Pages.

• push to main: build and deploy
• scheduled / manual workflows: refresh data, recompute heat, commit updated data, then deploy

This separation keeps normal code pushes fast while preserving a durable daily data pipeline.

npm install
npm run dev

Build locally:

npm run build
npm run preview

Run the data pipeline manually:

npm run fetch
python scripts/summarize.py
python scripts/pull_citation_signals.py
python scripts/pull_hf_signals.py
python scripts/compute_heat_scores.py
python scripts/export_heat_json.py

If you use AI summarization locally, set:

export GROQ_API_KEY="..."

src/
  components/        UI components
  config/            category, site, and feature config
  data/              core paper data and derived ranking artifacts
  layouts/           Astro layouts
  lib/               search, heat, grouping, i18n helpers
  pages/             bilingual routes
  styles/            global styles

scripts/
  fetch.js
  summarize.py
  pull_citation_signals.py
  pull_hf_signals.py
  compute_heat_scores.py
  export_heat_json.py

The project is actively iterating on:

• lighter default browse flows
• clearer ranking surfaces
• optional-depth search
• stronger bilingual consistency
• better visual hierarchy across home, latest, and rising views

MIT