RefBib

Overview

RefBib extracts references from academic PDFs and resolves them into standardized BibTeX entries — in one click. No LLMs, no hallucinated metadata. Every citation is verified against real academic databases: CrossRef, Semantic Scholar, DBLP, or GROBID's structured parsing.

Pipeline

1. Upload — Drop up to 20 PDFs with sequential batch processing. Append more at any stage without losing results.
2. Parse — GROBID extracts structured references from each PDF with multi-instance fallback.
3. Resolve — Waterfall resolution: CrossRef → Semantic Scholar → DBLP → GROBID fallback. Each entry gets a match status (Matched / Fuzzy / Unmatched).
4. Deduplicate — DOI, fingerprint, and bigram similarity detect duplicates across papers. Conflicts enter a resolution queue.
5. Export — Download deduplicated .bib or occurrence-preserving BibTeX. Copy-to-clipboard for quick use.

Key Features

• Workspace — Local-first reference management with automatic deduplication, conflict resolution, search, and status filtering.
• Analytics — Citation year distribution, venue breakdown, match quality donut chart, and most-cited references ranking.
• Manual Override — BibTeX editor for individual entries, DOI resolution for unmatched refs, Google Scholar links for every reference.
• Light/Dark Theme — Full theme support across all views.
• Self-Hostable — One-command setup (./start.sh), optional password protection, configurable GROBID instances.

Tech Stack

• Frontend — Next.js (App Router), shadcn/ui, Tailwind CSS, Recharts
• Backend — Python, FastAPI, httpx, lxml
• PDF Processing — GROBID with TEI XML output (local Docker or remote instances)
• Data Sources — CrossRef, Semantic Scholar, DBLP
• Deployment — Fly.io (backend) + Vercel (frontend)

My Role

Everything — parsing pipeline, waterfall resolution strategy, scoring heuristics, workspace UX, analytics visualizations, and deployment infrastructure. Built entirely through Claude Code.