citracer

📝 Description

A paper cites 50+ references, but which ones actually discuss the concept you care about? And which papers those cite? And the ones after that?

citracer answers this recursively. Give it a PDF and a keyword: it finds every sentence where the keyword appears, identifies the references cited nearby, downloads those papers, and repeats the process N levels deep. The output is an interactive citation graph you can explore in your browser. A 5-depth trace starting from a single paper typically surfaces 50-150 relevant papers in minutes.

With --semantic, matching goes beyond literal keywords: a sentence-transformer embedding model catches passages that express the same concept with different vocabulary (e.g. "univariate processing" matches a trace for "channel-independent").

With --reverse, citracer walks the other direction: "which papers cite this paper while mentioning a given concept?" This is useful for tracing how an idea spread forward through the literature.

Supported sources. citracer resolves cited papers through arXiv, Semantic Scholar, OpenReview, Sci-Hub, and Semantic Scholar's open-access PDF links (which cover PMC, publisher OA pages, and more). Preprint servers bioRxiv, medRxiv, ChemRxiv, SSRN, PsyArXiv, AgriXiv, and engrXiv are also supported. Papers that still can't be downloaded appear as unavailable nodes, but can be enriched with metadata via OpenAlex using the --enrich flag.

⚙️ Installation

Requirements: Python 3.10+ and Docker.

From PyPI (recommended)

pip install citracer
docker pull lfoppiano/grobid:0.9.0
docker run --rm -p 8070:8070 lfoppiano/grobid:0.9.0

# Optional: enable semantic matching (adds ~500MB for sentence-transformers + PyTorch)
pip install citracer[semantic]

After pip install citracer, the citracer command is available globally on your PATH. You can then run it from anywhere in your terminal:

citracer --pdf paper.pdf --keyword "your keyword"

From source (for development)

git clone https://github.com/marcpinet/citracer
cd citracer
pip install -e .
docker run --rm -p 8070:8070 lfoppiano/grobid:0.9.0

GROBID must be reachable on http://localhost:8070. Verify with curl http://localhost:8070/api/isalive.

1. --s2-api-key <key> as a CLI flag
2. S2_API_KEY environment variable in your shell
3. A persistent user config at ~/.citracer/config.json, set once via:

   citracer config set-s2-key <your-key>

Other config commands: citracer config show, citracer config get-s2-key (masked), citracer config clear-s2-key, citracer config path. The file is created with mode 600 on POSIX so other local users can't read it.

4. A .env file at the project root (copy .env.example and fill it in):

   S2_API_KEY=your_key_here
   

   The .env file is git-ignored.

If none of these are set, the unauthenticated public endpoint is used as fallback (much slower, frequent 429 backoffs).

Or pass it via `--email` or the `OPENALEX_EMAIL` environment variable.

## 🚀 Usage

After `pip install citracer` the `citracer` command is on your `PATH`. The examples below use it directly. If you cloned the repo instead, use `python -m citracer` in place of `citracer`.
```bash

# From a local PDF
citracer --pdf test_data/crossad.pdf --keyword "channel-independent" --depth 5

# From an arXiv id (auto-downloads the root PDF)
citracer --arxiv 2211.14730 --keyword "self-attention"

# From a DOI or URL (arXiv, OpenReview, bioRxiv, medRxiv, SSRN)
citracer --doi 10.48550/arxiv.2211.14730 --keyword "patching"
citracer --doi 10.1016/j.isci.2018.09.017 --keyword "attention"
citracer --url https://openreview.net/forum?id=cGDAkQo1C0p --keyword "instance normalization"
citracer --url https://www.biorxiv.org/content/10.1101/2024.01.01.123456v1 --keyword "CRISPR"

# Multi-keyword tracing (union by default, --match-mode all for intersection)
citracer --pdf paper.pdf --keyword "channel-independent" --keyword "patching"

# Reverse trace: find papers that cite the source while mentioning the keyword

# in their citation context. No PDF downloads, pure S2 metadata. Limit is optional.
citracer --arxiv 2211.14730 --keyword "channel-independent" --reverse --reverse-limit 500

# Enrich unavailable nodes with metadata (abstract, citation count) via OpenAlex
citracer --pdf paper.pdf --keyword "attention" --enrich --email [email protected]

# Supply a local PDF for a node that citracer couldn't download

# Supply a local PDF or a URL for a node that citracer couldn't download
citracer --pdf paper.pdf --keyword "attention" --supply-pdf "doi:10.1234/foo=~/papers/foo.pdf"
citracer --pdf paper.pdf --keyword "attention" --supply-pdf "title:abc123=https://example.com/paper.pdf"

# Export the graph for downstream analysis
citracer --pdf paper.pdf --keyword "..." --export out/graph.json --export out/graph.graphml

# Diff against a previous trace to highlight new papers (orange nodes)
citracer --pdf paper.pdf --keyword "attention" --diff output/old_graph.json
citracer --pdf paper.pdf --keyword "attention" --diff output/old_graph.json --since 2025
citracer --pdf paper.pdf --keyword "attention" --since 2025-06

# Re-run a trace without re-resolving papers already cached locally
citracer --pdf paper.pdf --keyword "attention" --depth 7 --no-refetch

# Semantic matching: catch conceptual matches the regex misses

# (requires: pip install citracer[semantic])
citracer --pdf paper.pdf --keyword "channel-independent" --semantic
citracer --pdf paper.pdf --keyword "attention" --semantic --semantic-threshold 0.55

Source (exactly one required)

Trace options

Output

🎨 Output

Nodes are colored by status:

Edges come in two flavors:

Interactive controls

A control panel in the top-left corner of the graph lets you tune the view on the fly:

Other interactive features:

• Hover any node → side panel updates live with title, authors, year, citation count, status, centrality metrics (PageRank, betweenness, in/out degree), a PIVOT badge for pivot papers, keyword hits (regex matches are highlighted; semantic matches show a purple SEM badge with the note "conceptual match") and a collapsible abstract section when available
• Search box in the control panel → fuzzy match by title or author, click a result to focus-and-pin the matching node
• Click a node → pins the panel; a blue border is drawn around the node to show the pinned state. The pin survives clicks on the empty canvas, hover on other nodes, and pan/zoom. It's only released by clicking the same node again, pressing the × close button on the info panel, or picking Unpin from the right-click menu
• Right-click any node → context menu with Hide, Pin/Unpin, Open link, and Supply PDF (on unavailable nodes only, prompts for a local path or URL and generates the --supply-pdf command to copy for the next run)
• Drag any node anywhere. After initial placement the layout is released, so nothing snaps back
• Undo / Redo with Ctrl+Z / Ctrl+Y (or Ctrl+Shift+Z). Reverts node positions, filters, layout, and all control settings. History of up to 50 snapshots
• Resizable panel: drag the right edge of the control panel to widen or narrow it
• show N more in a panel with many hits → expands the full list
• LaTeX math in passages is rendered with KaTeX ($...$, $$...$$, \(...\), \[...\])
• Automatic state persistence. Node positions, filters, pin state, dropdowns, spread slider and manually hidden nodes are all saved to localStorage keyed on a hash of the node-id set. A browser refresh restores the exact view you had. A new trace with a different paper set gets a fresh slate. The reset link at the bottom of the panel clears everything and reloads

Bibliometric analytics

Every trace automatically computes quantitative metrics on the citation graph:

The analytics collapsible section in the control panel shows global metrics, a clickable list of pivot papers (clicking focuses the node), and a keyword density timeline table with mini bar charts. Per-node metrics appear in the info panel when hovering or clicking a node.

All analytics are included in the JSON export ("analytics" key), the GraphML export (betweenness, pagerank, is_pivot as node attributes), and the reproducibility manifest.

Reproducibility

Every trace generates a manifest.json alongside the graph output, encoding everything needed to reproduce the exact same graph:

• citracer version, timestamp, full CLI command
• Source paper: type (pdf/doi/arxiv/url), raw input value, resolved title/DOI/arXiv ID
• Parameters: keywords, match mode, depth, context window, consolidate, reverse, enrich, GROBID URL
• Environment: Python version, platform, GROBID availability, API key/email status
• Results: node/edge counts, status breakdown, analytics summary (global metrics, timeline, pivot papers)

This allows anyone receiving a citracer graph to re-run the trace with identical settings. The manifest is also embedded in JSON exports under the "metadata" key.

🔍 How it works

1. PDF parsing. GROBID processes the PDF and returns TEI XML. citracer walks the <body> to reconstruct the plain text while recording the character offset of every inline <ref type="bibr"> citation. The bibliography is extracted from <listBibl>. Figure-diagram paragraphs (detected by their density of mathematical Unicode characters) are skipped to avoid polluting the keyword matcher. Paragraphs that GROBID splits mid-sentence around narrative citations (a common pattern around "Since Smith et al. (2020) and Jones et al. (2021) have shown...") are glued back together with a length-preserving regex so sentence-based matching still sees the refs and the keyword together.

2. Inline ref recovery. GROBID occasionally misses narrative citations like DLinear Zeng et al. (2023), especially when the author name isn't preceded by a parenthesis. A supplementary pass scans the text for canonical author-year patterns (Surname et al. (Year), Surname & Other (Year), Surname (Year)) and adds them as inline refs whenever the (surname, year) signature matches a unique bibliography entry. In typical ML papers this recovers dozens of refs per document.

3. Keyword matching (regex + optional semantic). The keyword is first compiled to a flexible regex that handles morphological variants (e.g. channel-independent matches channel-independence, channel independently, channelindependence). The body is segmented into sentences with pysbd, and each occurrence of the keyword is associated with the references cited in the same sentence or the immediately following one. When --semantic is enabled, a second pass scans every sentence the regex didn't match using a sentence-transformer embedding model: sentences whose embedding is close enough to the keyword (cosine similarity above the threshold) are added as additional hits. This catches conceptual matches where the idea is expressed with entirely different vocabulary (for example, tracing "channel-independent" also surfaces passages about "decoupled cross-channel correlations" or "per-variate processing"). Regex hits and semantic hits are unioned, so --semantic only adds recall without losing any existing matches.

4. Reference resolution. Each cited paper is resolved through the following cascade:

   1. If GROBID extracted a DOI or arXiv ID, use it directly.
   2. Otherwise, search arXiv by title (phrase first, then keyword fallback, with rapidfuzz validation). Search results are validated by fuzzy title match (threshold 85) and year cross-check (±3 years from the bibliography entry's year when known) to prevent false matches on similarly-titled papers from different eras.
   3. If arXiv has nothing, query Semantic Scholar with the same title + year validation, plus 429-aware backoff (also retrieves citation count and open-access PDF URL).
   4. As a last resort, search OpenReview (covers ICLR/TMLR papers not on arXiv).
   5. If --enrich is set, query OpenAlex for missing metadata (abstract, citation count, OA URL).

   PDF download cascade (in order): user-supplied PDF (--supply-pdf) > arXiv > OpenReview > Sci-Hub (by DOI, tries multiple mirrors) > S2 open-access URL (covers PMC, publisher OA, bioRxiv, medRxiv, etc.) > preprint-specific download (bioRxiv, medRxiv, ChemRxiv, SSRN, PsyArXiv, AgriXiv, engrXiv).

   All resolved PDFs and metadata are cached in ./cache/.

5. Recursion. The tracer is a BFS that processes papers in queue order. Each level's PDFs are parsed in parallel via a thread pool (--grobid-workers, default 4), and the reference resolves inside a single paper are also parallelized. Deduplication uses a canonical ID (DOI > arXiv > OpenReview > title hash). When the same PDF is reached via a second path, the new edge is added without re-parsing. Years from bibliography entries can backfill a node's year when older (e.g. a preprint v1 2022 takes precedence over a publication year 2023), but only within a ±2 year window of the first year we ever saw for that node. This prevents cascading from parser mistakes.

6. Cross-graph bibliographic links. After the recursive trace is complete, a post-processing pass scans every parsed paper's bibliography against every other node in the graph and adds dashed "bibliographic link" edges for pairs that cite each other but not in the keyword's neighborhood. Matching is exact on DOI/arXiv IDs and fuzzy (rapidfuzz, threshold 88) on titles. No external API calls are needed: everything runs on the already-in-memory graph, so the cost is negligible.

7. Bibliometric analytics. After the trace completes, citracer computes per-node centrality metrics (PageRank, betweenness) and graph-wide statistics (density, connected components, keyword density timeline) using networkx. Pivot papers (the earliest keyword-matched paper in each connected component, plus high-betweenness nodes with the keyword) are automatically flagged. A reproducibility manifest (manifest.json) is written alongside the graph, encoding the full trace parameters, environment, and results.

8. Rendering. The graph is serialized to an interactive HTML page using pyvis, with a custom overlay providing the layout/size/spread controls, the legend filters, the side info panel, keyword highlighting, and KaTeX math.

Reverse trace mode (--reverse)

The forward algorithm walks DOWN from a root paper into its bibliography. --reverse walks UP: "who cites this paper, and which of them mention the keyword in their citation context?".

The key trick is that Semantic Scholar's /paper/{id}/citations endpoint returns a contexts field for each citing paper: an array of 1-2 sentence snippets around every place that paper cites the source. We apply the same morphological keyword regex to those snippets locally. A paper whose citation contexts don't contain the keyword is rejected without downloading anything. A paper with a matching context is added to the graph with the snippet as its keyword_hits, plus its title/authors/year/arxiv-id from S2 metadata. No GROBID call, no arXiv download. (--semantic is not available in reverse mode because the snippets are too short for reliable embedding-based matching.)

For a paper with 2000+ citations, this runs in ~10-30 seconds and typically surfaces 20-100 relevant papers, depending on how specific the keyword is.

Semantic matching (--semantic)

The default regex handles morphological variants (e.g. channel-independent matches channel-independence, channel independently) but misses papers that express the same concept with different vocabulary: "univariate processing", "per-channel modeling", "decoupled channel correlations".

--semantic adds a second pass after the regex: every sentence the regex didn't already match is embedded with a sentence-transformer model (default: all-mpnet-base-v2, ~420MB) and compared to the keyword by cosine similarity. Sentences above the threshold (default 0.40) are added as additional hits. The result is a union: all regex matches plus any conceptual matches the embedding caught.

pip install citracer[semantic]
citracer --pdf paper.pdf --keyword "channel-independent" --semantic

Semantic hits appear in the info panel with a purple SEM badge labeled "conceptual match", so the user can distinguish them from regex hits at a glance. The header also shows a breakdown (e.g. "7 keyword hit(s) (5 regex + 2 semantic)").

--semantic-model NAME switches to a different model (e.g. all-MiniLM-L6-v2 for a lighter 80MB alternative). --semantic-threshold T tunes the similarity cutoff. Both flags imply --semantic.

Citracer can compare a new trace against a previous one to highlight what changed. This turns a one-shot snapshot into a monitoring workflow:

# Initial trace - export a baseline
citracer --pdf paper.pdf --keyword "attention" --depth 3 --export baseline.json

# Months later - re-run the same trace and diff
citracer --pdf paper.pdf --keyword "attention" --depth 3 --diff baseline.json

Nodes that weren't in baseline.json are colored orange and labeled NEW in the info panel. Their original status (analyzed, no match, unavailable) is preserved; the orange overlay is purely visual. The legend gains a clickable new (since last run) row to toggle them on/off.

--since YYYY or --since YYYY-MM highlights nodes published on or after a date. When combined with --diff, both conditions must be met (intersection): the paper must be absent from the baseline and published after the given date.

Month-level precision uses the publicationDate field from Semantic Scholar (YYYY-MM-DD) when available. When only the publication year is known, --since 2024-06 falls back to a year-only comparison (year >= 2024). Nodes with no known date are skipped and a warning is logged.

Both is_new flags (on nodes and edges) are included in JSON and GraphML exports, so downstream scripts can consume the diff without re-running citracer.

argparse entry point + GROBID health check + .env loader

├── pdf_parser.py

GROBID + TEI walking + figure-noise filter + paragraph merge + narrative ref supplementation + pymupdf fallback

├── keyword_matcher.py

morphological regex + sentence-based ref association (pysbd)

├── reference_resolver.py

arXiv-first cascade resolver (arxiv → S2 → OpenReview → Sci-Hub → OA → preprints) with SQLite cache

├── source_resolver.py

routes --pdf / --doi / --arxiv / --url inputs to a local PDF path

├── preprint_resolver.py

maps DOIs to preprint server PDF URLs (bioRxiv, medRxiv, ChemRxiv, SSRN, PsyArXiv, AgriXiv, engrXiv)

├── metadata_enrichment.py

OpenAlex API client for enriching nodes with abstract, citation count, and OA URLs

├── metadata_cache.py

SQLite-backed key/value store for resolver metadata, thread-safe

├── analytics.py

bibliometric metrics: PageRank, betweenness, pivot detection, timeline

├── cross_citation.py

post-trace pass that adds dashed bibliographic-only edges between graph nodes

├── diff.py

--diff / --since: compare against a previous trace, mark new nodes/edges

├── tracer.py

BFS recursion with parallel parsing, deduplication, year anchoring

├── visualizer.py

pyvis rendering pipeline

├── exporter.py

GraphML / JSON export (includes analytics and manifest)

├── manifest.py

reproducibility manifest generation

├── models.py

dataclasses

├── api_types.py

TypedDicts for arxiv / Semantic Scholar / OpenReview / OpenAlex payloads

├── constants.py

every tunable threshold and timeout, in one place

├── user_config.py

persistent user-level config (~/.citracer/config.json)

├── utils.py

ID normalization, hashing, tqdm-safe logging setup

└── templates/ └── overlay.html.tmpl

the interactive control panel (HTML/CSS/JS) injected into the pyvis output

## 🧩 Dependencies

<div class="table-wrapper">
<table>
<thead>
<tr><th><p>Package</p></th><th><p>Used for</p></th></tr>
</thead>
<tbody>
<tr><td><p><a href="https://github.com/kermitt2/grobid">GROBID</a></p></td><td><p>PDF structural parsing (external service)</p></td></tr>
<tr><td><p><a href="https://lxml.de/">lxml</a></p></td><td><p>TEI XML processing</p></td></tr>
<tr><td><p><a href="https://pymupdf.readthedocs.io/">pymupdf</a></p></td><td><p>PDF text extraction (parser fallback)</p></td></tr>
<tr><td><p><a href="https://github.com/lukasschwab/arxiv.py">arxiv</a></p></td><td><p>arXiv search and download</p></td></tr>
<tr><td><p><a href="https://github.com/nipunsadvilkar/pySBD">pysbd</a></p></td><td><p>Sentence boundary detection</p></td></tr>
<tr><td><p><a href="https://pyvis.readthedocs.io/">pyvis</a></p></td><td><p>Interactive HTML graph rendering</p></td></tr>
<tr><td><p><a href="https://github.com/rapidfuzz/RapidFuzz">rapidfuzz</a></p></td><td><p>Fuzzy title matching</p></td></tr>
<tr><td><p><a href="https://networkx.org/">networkx</a></p></td><td><p>Graph analytics (centrality, components, PageRank)</p></td></tr>
<tr><td><p><a href="https://requests.readthedocs.io/">requests</a></p></td><td><p>HTTP client</p></td></tr>
<tr><td><p><a href="https://github.com/tqdm/tqdm">tqdm</a></p></td><td><p>Progress bar</p></td></tr>
<tr><td><p><a href="https://github.com/theskumar/python-dotenv">python-dotenv</a></p></td><td><p>Loading the Semantic Scholar key from a <code>.env</code> file</p></td></tr>
<tr><td><p><a href="https://www.sbert.net/">sentence-transformers</a></p></td><td><p>Semantic matching via sentence embeddings (optional: <code>pip install citracer[semantic]</code>)</p></td></tr>
<tr><td><p><a href="https://katex.org">KaTeX</a></p></td><td><p>LaTeX math rendering in the HTML output (CDN)</p></td></tr>
<tr><td><p><a href="https://visjs.github.io/vis-network/docs/network/">vis-network</a></p></td><td><p>Interactive network rendering (via pyvis, CDN)</p></td></tr>
</tbody>
</table>
</div>

External APIs:

- [arXiv API](https://info.arxiv.org/help/api/index.html)
- [Semantic Scholar Graph API](https://api.semanticscholar.org/api-docs/graph)
- [OpenReview API](https://docs.openreview.net/reference/api-v2)
- [OpenAlex API](https://docs.openalex.org/) (metadata enrichment, opt-in via `--enrich`)
- [Sci-Hub](https://sci-hub.in/) (paywall bypass for PDF download)
- Preprint servers: [bioRxiv](https://www.biorxiv.org/), [medRxiv](https://www.medrxiv.org/), [ChemRxiv](https://chemrxiv.org/), [SSRN](https://www.ssrn.com/), [PsyArXiv](https://psyarxiv.com/), [AgriXiv](https://agrixiv.org/), [engrXiv](https://engrxiv.org/) (PDF download via DOI detection)

## ⚠️ Limitations

- GROBID misclassifies a small fraction of references, in particular sub-citations with letter suffixes like `Liu et al., 2024b`, which the supplementation pass can't disambiguate. These are silently dropped.
- The narrative-citation supplementation pass skips ambiguous `(surname, year)` signatures (e.g. two different Zhou 2022 papers in the bibliography). These missed cases are rare but do happen in survey-heavy papers.
- pysbd handles most academic abbreviations but can occasionally split mid-sentence; falling back to `--context-window 300` is sometimes useful.
- arXiv enforces ~3 seconds between requests, so the first run on a deep trace can take several minutes. The local cache makes subsequent runs fast.
- Papers that cannot be resolved through any source in the download cascade (arXiv, OpenReview, Sci-Hub, S2 open-access, preprint servers) appear as `unavailable` red nodes. Books and some workshop proceedings are typically not retrievable. Use `--supply-pdf` to provide PDFs manually (local path or URL) for these nodes.
- The "Fruchterman-Reingold" layout option is implemented via vis.js's `forceAtlas2Based` solver, which is the closest approximation available natively. A proper Kamada-Kawai implementation isn't offered because vis.js doesn't ship one.
- `--semantic` matching depends on the quality of the sentence-transformer model. The default `all-mpnet-base-v2` was benchmarked at F1=0.93 on academic citation text (vs 0.86 for `all-MiniLM-L6-v2`). Domain-specific keywords may benefit from threshold tuning. Semantic matching is not available in reverse trace mode.

## 🧪 Development
```bash
pip install -r requirements-dev.txt
pytest tests/ -v

The test suite is hermetic, with no GROBID and no network. GROBID output is exercised via a pre-baked TEI fixture in tests/fixtures/sample.tei.xml, and every external API (arXiv, Semantic Scholar, OpenReview, PDF downloads) is mocked. Runs in under 2 seconds.

CI runs the suite on Python 3.10 / 3.11 / 3.12 via GitHub Actions on every push to main, every pull request, and on manual dispatch from the Actions tab. See .github/workflows/tests.yml.

📖 Citation

If you use citracer in your research, please cite it as:

@software{pinet2026citracer,
  author       = {Pinet, Marc},
  title        = {citracer: Keyword and Concept-Driven Citation Graph Tracer},
  year         = {2026},
  url          = {https://github.com/marcpinet/citracer},
  note         = {Python package available at \url{https://pypi.org/project/citracer/}}
}

✍️ Authors

• Marc Pinet - Initial work - marcpinet