Daily Paper Digest Bot

This repository runs a GitHub Actions workflow that fetches recent arXiv papers, selects three papers per day, generates an English research brief with Gemini, and sends the digest by email.

What It Does

Each daily run executes this pipeline:

1. Fetch recent arXiv candidates from robotics- and ML-related queries.
2. Rank papers into three buckets:
   • most relevant to your recent work
   • interest-based topic recommendation
   • a representative paper from a configured author shortlist in the last three months
3. Generate an English digest with Gemini.
4. Send the digest through SMTP.
5. Persist sent paper IDs to avoid repeats across future runs.

If Gemini summarization fails, the pipeline automatically falls back to a plain English template email built directly from the selected paper metadata.

Repository Layout

.
├── .github/workflows/daily_papers.yml
├── config/topics.yaml
├── data/history.json
├── requirements.txt
└── scripts
    ├── fetch_arxiv.py
    ├── main.py
    ├── rank_papers.py
    ├── send_email.py
    └── summarize_with_llm.py

Requirements

• Python 3.11
• A Gemini API key (Gemini has free quota every month. You can use your own API keys from other sources, but you need to modify [summarize_with_llm.py].)
• An SMTP account that supports STARTTLS
• A GitHub repository with Actions enabled

GitHub Actions Deployment

The workflow file is .github/workflows/daily_papers.yml. Currently it is scheduled to run every day at 00:30 UTC, which is 08:30 in Beijing/Taipei.

To deploy:

1. Push this repository to GitHub.
2. In the repository, open Settings -> Secrets and variables -> Actions.
3. Add these repository secrets:
   • GEMINI_API_KEY
   • SMTP_HOST
   • SMTP_PORT
   • SMTP_USER
   • SMTP_PASS
   • EMAIL_TO
4. Optionally add GEMINI_MODEL if you want to override the default model. If unset, the code uses gemini-2.5-flash.
5. In Settings -> Actions -> General, make sure GitHub Actions is enabled.
6. In the workflow permissions section, allow read and write access to repository contents, because the workflow commits data/history.json and data/latest_digest.json.
7. Manually trigger the workflow once from the Actions tab with workflow_dispatch to verify the setup.

SMTP Notes

The email sender uses SMTP with STARTTLS. Typical values look like:

• SMTP_HOST: smtp.gmail.com. (If you are using Gmail)
• SMTP_PORT: 587

If you use Gmail, you will usually need an app password rather than your normal account password.

Local Setup

Create and activate a Conda environment, then install dependencies:

conda create -n paper-digest python=3.11 -y
conda activate paper-digest
pip install -r requirements.txt

Set the required environment variables:

export GEMINI_API_KEY="..."
export GEMINI_MODEL="gemini-2.5-flash"
export SMTP_HOST="smtp.example.com"
export SMTP_PORT="587"
export SMTP_USER="bot@example.com"
export SMTP_PASS="..."
export EMAIL_TO="you@example.com"

Run the pipeline:

python scripts/main.py

Configuration

Edit config/topics.yaml to tune:

• positive keywords for recent-work matching
• negative keywords to suppress irrelevant domains
• topic interests
• spotlight authors and their lookback window
• arXiv source queries
• the Gemini model through the GEMINI_MODEL environment variable if needed

Outputs

Each successful run updates:

• data/history.json: sent paper IDs used for deduplication
• data/latest_digest.json: the three selected papers from the latest run

Operational Notes

• The workflow depends on external services: arXiv, Gemini, and your SMTP provider.
• GitHub Actions scheduled workflows are not guaranteed to start at the exact minute.
• If no eligible papers are selected, the script exits without sending email.
• The repository currently stores deduplication state in Git. That is simple and effective for a single daily workflow, but it is not designed for high-frequency concurrent runs.