AI Papers Reader is an AI agent that brings you weekly digests of latest AI papers, customizable to topics you care about. Check out the published digests at https://ai-papers-reader.taodong.net/.
AI Papers Reader is built with the following building blocks:
- Hugging Face's Daily Papers API: It's used to retrieve the metadata of recently published AI papers.
- Google Gemini, DeepSeek, or Claude: The selected LLM processes paper metadata, identifies papers relevant to the topics in the prompt, and summarizes recommended papers.
- Github Actions: A workflow runs automatically on Fridays to retrieve the latest paper metadata and use the AI model to generate a new digest. The digests are saved in the docs folder as markdown files.
- Netlify: The markdown files are then deployed to a static website using Netlify.
The default set of topics AI Papers Reader currently uses to identify relevant papers is based on my research interests. You can customize it by forking the repo and editing the prompt template.
Provider and model selection live in the committed config.yaml. The supported provider defaults are:
| Provider | Default model | Credentials |
|---|---|---|
| Gemini | gemini-flash-latest |
GOOGLE_API_KEY |
| DeepSeek | deepseek-v4-flash |
DEEPSEEK_API_KEY |
| Claude | claude-haiku-5-5 |
ANTHROPIC_API_KEY or ant auth login locally; Workload Identity Federation in GitHub Actions |
The current committed selection is Claude:
provider: claude
model: claude-haiku-5-5To use Gemini or DeepSeek instead, change those two values to one of:
provider: gemini
model: gemini-flash-latestprovider: deepseek
model: deepseek-v4-flashStore credentials locally in an ignored .env file. Start from .env.example:
cp .env.example .env
# Edit .env and set GOOGLE_API_KEY, DEEPSEEK_API_KEY, and/or ANTHROPIC_API_KEY.The CLI --provider and --model flags remain available as one-off overrides. If no model override is supplied, the configured model is used when it matches the selected provider; otherwise the provider default from the table above is used.
pip install -r requirements.txt
python src/generate_report.pyDeepSeek PDF summarization extracts text locally, so scanned PDFs without a text layer are not supported.
PDF handling is bounded for reliability: Gemini and DeepSeek receive the complete paper when its PDF is under the configured 15 MiB default. Claude, which reads the PDF itself at about 3,000 tokens per page, receives only the main text: each PDF ends on the page with its References/Bibliography heading. Dropping the reference list and appendices typically removes about two-thirds of a paper's pages and keeps most requests under Claude Haiku 5.5's 100K-token price tier. For oversized PDFs, text extraction/PDF input first tries the complete body before a configurable References/Bibliography heading, then drops trailing pages one at a time from the configured page limit until the generated PDF fits. Downloads stream in chunks and retry transient network/HTTP failures with exponential backoff; tune the PDF timeout/retry settings in config.yaml as needed. Set pdf.stop_at_references to false or customize pdf.references_headings when needed.
For GitHub Actions, commit the desired config.yaml selection. Gemini and DeepSeek read their API key from a repository secret (GOOGLE_API_KEY or DEEPSEEK_API_KEY).
Claude needs no secret: both workflows authenticate through Workload Identity Federation. They grant id-token: write and set the non-secret ANTHROPIC_FEDERATION_RULE_ID, ANTHROPIC_ORGANIZATION_ID, ANTHROPIC_SERVICE_ACCOUNT_ID, and ANTHROPIC_WORKSPACE_ID IDs. GitHub's OIDC tokens expire about five minutes after issue and can each be exchanged only once, so src/llm_client.py requests a fresh token from GitHub whenever the SDK renews its short-lived Anthropic access token. The federation rule must match the run's GitHub token: scheduled runs, and manual runs started from main, present the subject repo:<owner>/<repo>:ref:refs/heads/main.