A personal arXiv radar that runs on GitHub's dime

What it does

Daily Paper Reader is a self-hosted academic paper platform that scrapes arXiv, OpenReview, and several bio/chem preprint servers daily, then surfaces personalized recommendations through a web UI. It bundles reading, summarization, and LLM-based Q&A into one flow, with Zotero integration for saving papers.

The entire backend runs on GitHub Actions: scheduled workflows fetch papers, generate embeddings, rerank candidates, and produce structured summaries. The frontend deploys to GitHub Pages. You bring your own LLM API key (DeepSeek by default) and a GitHub PAT; the project handles the rest.

The interesting bit

The architecture treats GitHub’s free tier as the infrastructure layer. The “dual pipeline” workflow (shown in the diagram) splits into a daily ingestion path and an on-demand reading path, both executing in CI runners. A recent pivot to pure frontend architecture moved subscription management and LLM calls into the browser, reducing server-side state.

Key highlights

• Multi-source ingestion: arXiv, OpenReview, bioRxiv, medRxiv, ChemRxiv, plus conference proceedings
• Hybrid retrieval: BM25 exact search, vector ANN with density fallback, and reranking (local Qwen3-0.6B or remote)
• Stateful recommendations: per-tag carryover windows and seen-paper tracking to avoid cross-contamination between research interests
• AI reading companion: inline Q&A with conversation history, plus auto-generated briefs and scored summaries
• Local debug mode: can run the full pipeline on CPU with LAN-accessible backend, useful for testing before pushing to Actions

Caveats

• The project is actively evolving with frequent breaking changes to configuration formats (recently removed Java/pdffigures2, old “Plato/BLT” model chains, etc.)
• Embedding and reranking services default to a third-party endpoint (zwwen.online) with unclear operational longevity; local fallback exists but requires setup
• Conference paper filtering is opinionated: only papers scoring 4+ are retained, which may exclude borderline work you care about

Verdict

Worth a look for researchers who want a customizable, code-first alternative to commercial paper feeds like Semantic Scholar or Mendeley. Skip it if you need stability over features—the changelog reads like a live refactoring exercise.