Discover the evolution of AI/ML research through interactive citation networks
👉 Try it now at https://paperverse.co
This tool helps you visualize and explore the interconnected world of Machine Learning research papers from the past decade, allowing you to:
🔍 Discover the most influential ideas in machine learning
🏗️ Trace the stepping stones of current state-of-the-art research
📈 Identify emerging trends in scholarly circles
Explore papers in a BFS (Breadth-First Search) manner to see how ideas evolve over time. Here's the citation graph starting from the landmark "Attention is all you need" paper:
Hover over any paper to get quick insights without diving deep:
Build your research roadmap by saving interesting papers for later:
Click on any paper to discover its most influential derivatives. Perfect for:
- Tracking idea evolution
- Finding research gaps
- Understanding paper lineage
┌────────────────────────────┐
│ Neo4j (Community + GDS) │
│ shared with ResearchQuest │
└──────────────┬──────────────┘
│ Cypher, bolt+s
┌──────────────┴──────────────┐
│ citation-network-backend │
│ (FastAPI) │
└──────────────┬──────────────┘
│ /api/*
┌──────────────┴──────────────┐
│ citation-network-ui │
│ (React, served by nginx) │
│ Firebase — auth + reading │
│ list only, not the graph │
└─────────────────────────────┘
Paperverse and ResearchQuest both explore the same underlying citation graph and now share one Neo4j instance instead of each maintaining a separate copy of the data. The backend here runs Cypher queries directly against it — no local database file, nothing to download or keep in sync by hand.
Paper ids throughout are Semantic Scholar's own paper ids (opaque hash strings), not sequential integers — this is a Neo4j-native identifier, a change from this project's original SQLite-backed version.
citation-network-backend/— FastAPI backend, Cypher queries over Neo4jcitation-network-ui/— React frontend (search, graph explorer, path finder, reading list); Firebase handles auth and the reading list only — none of the citation-graph data goes through Firebasebuild_graph/— the original arXiv → Semantic Scholar → graph-export pipeline this project was built around (see Known constraints below for where this is headed)docker-compose.yml/docker-compose.prod.yml/docker/— local dev and shared-VM deployment (see Deployment)
cd citation-network-backend
pip install -r requirements.txt
cp ../.env.example ../.env # fill in NEO4J_URI / NEO4J_USER / NEO4J_PASSWORD
uvicorn app:app --host 0.0.0.0 --port 8000 --reloadPoint NEO4J_URI at any Neo4j instance that already has the :Paper
graph loaded — see ResearchQuest's docker-compose.yml for a self-hosted
Neo4j + GDS setup, or run your own via build_graph/.
cd citation-network-ui
npm install
cp .env.example .env # Firebase config — see below
npm startFirebase (REACT_APP_FIREBASE_* in .env) is only used for login and the
reading-list feature. The app degrades gracefully without it — search,
graph exploration, and path finding all work with no Firebase config at
all (see citation-network-ui/src/firebase.js).
Runs alongside ResearchQuest on the same Hetzner VM, as a separate
docker compose project (docker-compose.prod.yml) — not its own server.
Two containers, frontend and backend, with no host ports published at
all: ResearchQuest's own nginx (already TLS-terminating Neo4j's bolt
connection for its Streamlit app) reaches Paperverse's frontend container
directly over a shared Docker network, and adds a second HTTPS server block
for paperverse.co alongside its existing one.
The domain itself is registered via AWS but its authoritative nameservers
are Cloudflare's — DNS changes happen in the Cloudflare dashboard, not
Route53 (a leftover, non-authoritative hosted zone exists there from an
earlier AWS-only deployment and can be ignored). TLS certificates are
issued by Let's Encrypt via certbot --webroot, auto-renewing through the
same mechanism ResearchQuest's own nip.io endpoint uses.
The VM's own address isn't published here — Neo4j's bolt endpoint has to stay reachable from Streamlit Cloud's servers (which have no fixed egress IPs to allowlist), so that port has to stay open to the internet, and there's no reason to make it an easy target for scanners that scrape public repos. Auth + TLS are the real protection either way.
- Citation data originates from Semantic Scholar's API, which restricts bulk redistribution of derived data. A meaningful fraction of papers in the graph — specifically ones that entered it only through citation expansion, rather than being one of the originally-fetched arXiv papers — don't have a title/abstract/publish-date match in Semantic Scholar's arXiv metadata, so fields like TL;DR and published date show as unavailable for them. A migration to OpenAlex (CC0-licensed, indexes arXiv ids natively, ships as a bulk downloadable snapshot rather than a rate-limited API) is the leading candidate to close this gap and also enable publishing a derived dataset publicly — not yet done.
- Very recent papers naturally have few incoming citation edges yet — a property of citation graphs generally, not a data quality bug.
The build_graph directory contains tools to create citation networks through these steps:
- Data Collection: Fetch ML papers from Arxiv (last 10 years)
- Citation Analysis: Gather citation data from Semantic Scholar
- Graph Generation: Create DOT files with papers as nodes and citations as edges
- Ranking: Apply PageRank to identify influential papers
- Database Integration: Process and export graph data for Neo4j