Citation Graph — Free ResearchRabbit Analog

Overview

A free, subscription-free alternative to ResearchRabbit and Connected Papers. Given a seed DOI (or several seeds), this skill walks the OpenAlex citation network outward in both directions and assembles a citation / co-citation graph of the local literature:

• Backward edges — the works each seed cites (its referenced_works).
• Forward edges — the works that cite each seed (its cited_by set).

It then ranks the discovered neighbourhood by co-citation strength — how many of the seed papers a candidate work connects to, via shared references, shared citers, and direct seed links — surfacing the papers most central to the topic, just like ResearchRabbit's "Similar Work" panels. The whole graph is exported as GraphML (open in Gephi, Cytoscape, yEd, or networkx) and JSON (for downstream code).

OpenAlex data is open (CC0) and the API is free for this scale of use, but since February 2026 requests are metered against a daily budget and the old mailto "polite pool" is gone. Keyless calls share $0.10/day per IP address — on a campus or cloud network that budget is often already spent, which surfaces as HTTP 429. A free API key (OPENALEX_API_KEY, from openalex.org/settings/api) gives $1/day of your own. A typical walk costs a fraction of a cent: seed lookups are free and each list/filter call costs $0.0001.

When to Use This Skill

• Map a literature landscape from a starting paper — "I have one key paper, show me the

field around it."

• Find seminal / highly co-cited papers in an area without manually chasing references.
• Snowball a reference network (forward and backward citation chasing) for a review.
• Build a citation map or run a co-citation analysis for a methods/visualization figure.
• Visualize how a research area connects — export to Gephi/Cytoscape for a network figure.

This is the graph-building / discovery tool; it needs a seed work to walk from.

Does NOT Trigger

Quick Start

export OPENALEX_API_KEY=...   # free at openalex.org/settings/api; recommended

# One seed DOI, default 1-hop walk, write both formats next to a basename:
uv run python scripts/build_graph.py \
    --seed 10.1038/nphys1170 \
    --out graph/seed1
# -> graph/seed1.graphml  +  graph/seed1.json

# Several seeds (DOI / OpenAlex W-id / arXiv id), deeper walk, larger ranking:
uv run python scripts/build_graph.py \
    --seed 10.1038/nphys1170 \
    --seed W2741809807 \
    --seed arXiv:2310.06825 \
    --depth 2 --per-seed 50 --top 40 \
    --out graph/transformer

# Offline smoke test (no network) — verifies the pipeline end-to-end:
uv run python scripts/build_graph.py --self-test

How It Works

1. Resolve seeds. Each --seed is normalized to an OpenAlex selector. Accepted forms:

a bare DOI (10.1038/nphys1170), a DOI URL, an OpenAlex work id (W2741809807), an OpenAlex URL, or an arXiv id (arXiv:1706.03762 / 1706.03762).

1. Expand backward. From each seed's referenced_works, add seed -> reference edges.
2. Expand forward. Query filter=cites:<id> to find works that cite the seed, adding

citer -> seed edges (capped at --per-seed; OpenAlex returns at most 100 per page).

1. Walk deeper. With --depth N, repeat the expansion on first-hop neighbours (N >= 2).
2. Rank by co-citation. Each non-seed node is scored by `shared_refs + shared_citers +

direct_seed_links — bibliographic coupling and co-citation against the seed set — with global cited_by_count` as the tie-breaker so canonical works float to the top.

1. Hydrate. References arrive as bare W-ids, so the ranked works' titles, years, and DOIs

are fetched in batched filter=openalex:W…|W… calls (100 ids per call) and the table is re-ranked.

1. Export. Write <out>.graphml and <out>.json.

Options

Output

GraphML (<out>.graphml) — a directed graph with node attributes title, year, doi, role (seed / reference / citation / neighbor), and cited_by_count. Open it directly in Gephi, Cytoscape, yEd, or networkx.read_graphml.

JSON (<out>.json) — schema alterlab-citation-graph/1.0:

{
  "schema": "alterlab-citation-graph/1.0",
  "seeds": ["10.1038/nphys1170"],
  "node_count": 142,
  "edge_count": 318,
  "nodes": [ { "id": "W…", "title": "…", "year": 2017, "doi": "…", "role": "citation", "cited_by_count": 8123 } ],
  "edges": [ { "source": "W…", "target": "W…" } ],
  "cocitation_ranking": [
    { "id": "W…", "title": "…", "cocitation": 4, "shared_refs": 2,
      "shared_citers": 1, "direct_seed_links": 1, "cited_by_count": 8123 }
  ]
}

Access & Limits

• Use a free key. Ask the user to set OPENALEX_API_KEY (openalex.org/settings/api) rather

than pasting the key into chat or a committed file. Without one the run still works when the shared per-IP budget has room, and the script prints a note.

• HTTP 429 means the daily budget is spent (it resets at midnight UTC) or the run exceeded

100 requests/s. The script stops with an explicit error instead of writing a silently truncated graph; add a key or retry later.

• Bound the walk. --depth 2 with a large --per-seed fans out fast (one list call per

frontier work); keep --per-seed reasonable and raise --sleep if you hit rate limits.

• Coverage caveat. OpenAlex citation coverage is broad but imperfect; very new

preprints may have sparse cited_by sets.

• arXiv seeds. An arXiv:<id> seed is resolved via its DataCite DOI

(10.48550/arXiv.<id>). If a preprint was later merged into its published-version record, OpenAlex may carry only the publisher DOI — pass that DOI as the seed instead.

Complementary Tools

Part of the AlterLab Academic Skills suite.