Citation Network Explorer

Interactive citation network explorer powered by live OpenAlex data, GPU-accelerated visualisation, and a built-in AI assistant.

Live demo: https://uoa-eresearch.github.io/citations/

Features

• Live OpenAlex search — type in the search box to load the citation network of any author, paper (the paper + its references + citing papers), institution, topic, journal or funder. Up to 600 works are loaded per entity, most cited first.
• Five views, switchable from the top bar (both network views run on cosmos.gl, the GPU force layout & rendering engine behind Cosmograph):
  • 3D — GPU 3D force simulation with an orbit camera (drag to rotate, scroll to zoom), shaded sphere points with depth cueing. Uses cosmos.gl's spaceDimensions: 3 mode, vendored from a pre-release build in vendor/cosmos-gl/ (see its README to rebuild/upgrade).
  • 2D — GPU 2D force layout that scales to very large graphs
  • Map — world map of author institutions and co-authorship collaborations (deck.gl arcs + scatter over a MapLibre/CARTO basemap)
  • Canvas — task-adaptive multi-axis 3D scatter (the dynamic canvas): assign contextual dimensions to each axis — time, citations, field-weighted impact, topic / subfield / field (ordered by the OpenAlex ontology hierarchy domain → field → subfield → topic), country, venue — or pick a task preset (Ideation over time, Geographic story, Impact landscape, Semantic map) and the axes reconfigure to suit the task. The AI assistant can also reconfigure the canvas from natural language, e.g. “where are the authors from and how has that shifted over time?”
  • Plots — Plotly analytics dashboard (works & citations per year, top authors / journals / topics / fields, open-access status, work types)
• Network styling — group nodes by paper, author, institution, journal, topic or field; colour and size by citations, year, author count, institution count or field.
• Timeline filter — drag across the year histogram to filter every view to a year range.
• AI assistant (✦ Ask AI) — ask questions by typing or voice (Web Speech API), e.g. “show Giovanni Coco's network on a map”, “plot citations per year”, “what's the most cited paper here?”, “which institutions does this author collaborate with?”, “what are the differing scholarly opinions around this work?” (it samples abstracts of citing papers and synthesises the discourse). The assistant uses Claude with tool use to drive the app (load networks, switch views, reconfigure the canvas axes, restyle, filter, highlight plots) and to answer data questions via live OpenAlex queries — including aggregations the app doesn't chart. Optionally speaks its replies.

AI setup

No setup needed on the CER network: if ai.cer-sandbox.cloud.edu.au is reachable, the app detects it on load and uses it automatically (its first advertised model), so the assistant just works. Otherwise it falls back to the Anthropic API and waits for you to add credentials.

To choose a different backend, open ✦ Ask AI → ⚙ settings and pick a provider. All credentials are stored only in your browser's localStorage and requests go directly from the browser to the provider (no server in between). Picking a provider yourself is remembered and disables the sandbox auto-detection. With no provider available, only simple typed commands work (“show map”, “load author name”).

• Anthropic API (default) — paste a Claude API key from console.anthropic.com. Model defaults to Claude Opus 5; server-side refusal fallback is enabled for Opus 5-class models.
• Amazon Bedrock — AWS region + credentials (access key / secret / optional session token) with bedrock:InvokeModel permission; requests are SigV4-signed in the browser straight to the Bedrock runtime (which supports CORS). Prefer short-lived STS/Cognito credentials. Model ID e.g. anthropic.claude-opus-5.
• Azure AI Foundry — your resource endpoint (e.g. my-resource.azure.anthropic.com) and API key; model e.g. claude-opus-5. If the browser blocks requests, enable CORS for your origin on the resource.
• OpenAI-compatible — the CER sandbox (https://ai.cer-sandbox.cloud.edu.au/v1, the default when reachable) or any other OpenAI-compatible server (Ollama, LM Studio, vLLM, …): base URL, optional key, and a model that supports tool calling — the model box autocompletes from the server's /v1/models. Tool-use quality varies by model. For Ollama served on a different origin, set OLLAMA_ORIGINS.

Running locally

The app is a static site (ES modules — it needs an HTTP server, not file://):

python3 -m http.server
# then open http://localhost:8000/

URL parameters

• ?data=Dan — load a pre-fetched local dataset from data/Dan.json
• ?type=author&id=A5043686349&name=Giovanni%20Coco — deep-link a live OpenAlex entity

Pre-fetched sample datasets

./fetch_author.py AUTHOR_NAME and ./fetch_paper.py PAPER_ID (after pip3 install -r requirements.txt) save datasets into data/ for offline use:

• Daniel J. Exeter (default)
• Giovanni Coco
• Mark Gahegan
• Quinn Asena
• Decolonizing Methodologies (map, 3D map)