📄 Airpush

From hundreds of papers to a daily top list — automatically.

Automated research scouting pipeline for labs and R&D teams.
Fetch, deduplicate, score, deeply analyze, and push key papers to Notion / WeChat.

arXiv · PubMed · RSS · Conference → Semantic Dedup → Rule Scoring → 3-Stage LLM Funnel → Daily Digest

English | 简体中文

---

What is Airpush?

Airpush is an automated paper intelligence workflow that turns a noisy daily paper stream into a compact, high-value digest.

Assign a domain config, run the pipeline, and receive:

• top-ranked papers with structured summaries,
• key figure extraction from PDFs,
• Notion-ready entries and markdown reports,
• optional WeChat / webhook notifications.

No manual triage, no repetitive filtering, and no copy-pasting paper notes.

---

Features

• Multi-source ingestion

  • arXiv category + query based crawling
  • PubMed (NCBI E-utilities)
  • RSS feeds (Nature, IEEE, etc.)
  • Conference sources (Semantic Scholar with OpenReview fallback)

• Three-stage filtering funnel

  • semantic dedup (all-MiniLM-L6-v2, cosine threshold)
  • rule-based scoring (institution whitelist + keyword matrix)
  • fast LLM relevance gate ({"relevant": true/false})

• Deep structured analysis

  • extracts ~20 research fields (task, modality, architecture, novelty, gains, limitations, reproducibility hints, etc.)

• Storage + delivery

  • SQLite local persistence (domain isolation)
  • Notion database + native block rendering
  • markdown daily report export
  • WeChat ServerChan / generic webhook push

• Figure intelligence from PDF

  • detect candidate figures automatically
  • rank/select, crop, host via GitHub + jsDelivr
  • embed to Notion with EN/CN caption support

---

Built-in Domains

Domain	Focus
medical	Medical imaging AI (diffusion, multimodal LLM, reconstruction, super-resolution)
cqed_plasmonics	Cavity QED & plasmonics (strong coupling, Purcell, nanophotonics)

Domain logic is decoupled through YAML. You can add new research directions by extending config files.

---

Quick Install

# Replace with your fork or the actual repository URL
git clone https://github.com/your-username/medpaper-flow.git
cd medpaper-flow
pip install -r requirements.txt

---

Getting Started

1) Configure runtime

cp .env.example .env

Edit one of:

• config/domains/medical/runtime.yaml
• config/domains/cqed_plasmonics/runtime.yaml

Minimal example:

llm:
  api_key: "your-api-key"         # use EMPTY for local vLLM if applicable
  base_url: "http://localhost:8000/v1"
  quick_model: "your-model-name"
  deep_model: "your-model-name"

run:
  dry_run: true

2) Dry run

DRY_RUN=true python main.py
DOMAIN=medical DRY_RUN=true python main.py

3) Production run

python main.py

---

Figure Extraction (Optional)

Enable in runtime.yaml:

llm:
  figure_hosting_enabled: true
  figure_selection_use_llm: true
  figure_max_images: 3
  github_token: "${GITHUB_TOKEN}"
  github_user: "your-username"
  github_repo: "paper-figures"
  github_branch: "main"

Security tip: keep secrets (API keys/tokens/webhooks) in environment variables instead of committing plaintext values.

---

Architecture (Pipeline)

Candidate Papers (arXiv / PubMed / RSS / Conference)
                    ↓
         Semantic Dedup + Rule Scoring
                    ↓
          LLM Quick Relevance Filter
                    ↓
        LLM Deep Structured Extraction
                    ↓
      PDF Download + Fulltext/Figure Analysis
                    ↓
  SQLite / Notion / Markdown / WeChat / Webhook

---

Configuration

Primary config path:

• config/domains/<domain>/runtime.yaml

Environment variables can override YAML values at runtime (for example via DOMAIN=medical).

---

Development

Run pipeline entrypoint:

python main.py

Core modules are under src/:

• fetchers: src/fetchers/
• ranking/filtering: src/scorer.py, src/deduplicator.py
• LLM pipeline: src/llm_pipeline.py
• persistence/push: src/storage/, src/notifier.py

---

License

MIT. See LICENSE.

---

简体中文

📄 Airpush

每天数百篇论文，自动筛出最值得读的少量论文。

面向科研团队 / 实验室的自动化论文情报管线。
自动采集、去重、评分、深度分析，并推送到 Notion / 微信。

arXiv · PubMed · RSS · Conference → 语义去重 → 规则评分 → LLM 三级漏斗 → 每日精选

English | 简体中文

Airpush 是什么？

Airpush 用于把“高噪声论文流”变成“低噪声高价值日报”。

配置好领域后，一次运行即可得到：

• 入选论文的结构化解读；
• PDF 关键图表提取；
• Notion 数据库同步与 Markdown 日报；
• 可选微信 / Webhook 通知。

核心能力

• 多源采集：arXiv、PubMed、RSS、会议源。
• 三级过滤：语义去重 + 规则评分 + LLM 快速相关性判断。
• 深度精析：抽取约 20 个结构化字段（任务、模态、架构、创新、局限等）。
• 存储与推送：SQLite、Notion 原生 Block、微信 ServerChan、Webhook。
• 图表智能提取：自动识别候选图，裁切上传并嵌入 Notion（支持中英文图注）。

内置领域

领域	方向
medical	医学影像 AI：扩散、多模态 LLM、重建、超分辨率
cqed_plasmonics	腔量子电动力学与等离激元：强耦合、Purcell、纳米光子

快速开始

git clone https://github.com/your-username/medpaper-flow.git
cd medpaper-flow
pip install -r requirements.txt
cp .env.example .env
DRY_RUN=true python main.py

配置文件

• 主配置：config/domains/<domain>/runtime.yaml
• 支持通过环境变量覆盖（例如 DOMAIN=medical）

图表功能开关（可选）

llm:
  figure_hosting_enabled: true
  figure_selection_use_llm: true
  figure_max_images: 3

开发入口

python main.py

许可证：MIT（见 LICENSE）。