S3 maps the themes of a research field from the titles and abstracts of its papers, checks which map fits your data best, and writes a report organised for a literature review. Use it from Claude, ChatGPT or a Google Colab notebook without writing code yourself, or from R.
The whole guide in 3 minutes. Watch on YouTube.
Select all results → Export → CSV, and tick every information group. Step by step, with what each field unlocks: Scopus export for S3.
Each extra field unlocks more of the analysis: references give citation flows and the map comparison, authors and affiliations give the who's who. Web of Science and OpenAlex exports also work.
Give S3 the file exactly as Scopus exported it. Opening and re-saving it in Excel cuts long reference lists at 32,767 characters.
uv starts S3 and installs everything it needs. Paste one line, then close and reopen the window.
With a key from dev.elsevier.com, S3 fetches every paper's full reference list from Scopus using the EID column of your export: exact citations, and bibliographic coupling from Scopus IDs instead of reference text. Without a key, S3 matches titles in the References column.
--scopus to the report command.Reference lists need your institution's Scopus access: run it on campus or the VPN, or ask your library for an institutional token. Never paste a key into a chat.
S3 uses the text map by default, as in the published method. If you ask for a hybrid map, S3 first screens text plus bibliographic coupling, co-citation or direct citation at 5%, 10%, … 40% weight. It uses the best hybrid only when it raises both the silhouette and the QI of the communities above the text map's, while clustering at least 90% as many papers; otherwise it keeps the text map. Citations only count for papers that are already close in content.
The report's Q0 shows the screening: silhouette, QI and topic coherence at every weight, and which map was used.
S3 can add what each paper uses and finds to its knowledge graph as triples: methods, data sources, theories, constructs, countries, industries and level of analysis, and findings between constructs such as absorptive capacity → increases → innovation performance. The report then shows each theme's methods and theories, gaps, and which findings papers agree or disagree on (Q39 and Q40).
Upload your own triples (CSV or Excel with paper_id, subject, relation, object), or let Claude or ChatGPT extract them from the abstracts. A progress bar shows each batch, and every triple keeps a quote from the abstract that S3 checks.
--triples ~/Desktop/triples.csv to the report command. --backend codex uses your ChatGPT plan.pip install "s3-bibliometric[llm]" and ANTHROPIC_API_KEY or OPENAI_API_KEY (Colab: Secrets).Try 30 papers first: extraction takes a few seconds per paper, and the answers are saved, so a stopped run carries on where it left off.
Tools missing? Run which uvx (Mac) or where uvx (Windows) and put that full path in "command". Updating: restart Claude; if the new tools don't show, run uv cache clean s3-bibliometric first.
Claude asks which language model to use (mpnet, from the S3 paper, is the default) and whether you have a Scopus API key. Never paste a key into the chat.
Colab runs on Google's servers, so your file is uploaded there. Scopus API key: store it under Secrets (key icon) as SCOPUS_API_KEY, then run import os; from google.colab import userdata, os.environ["SCOPUS_API_KEY"] = userdata.get("SCOPUS_API_KEY") and s3bib.enrich_citations(res) before the report; Google's servers are off campus, so this usually also needs an institutional token. In Jupyter on your computer, skip the upload cell and use the file path.
n if you have no Scopus API key. Add --map hybrid for a hybrid map (step 4).s3_report.html and click the downloads at the top: themes table, methods paragraph, reading list.Browser chats can't start programs on your computer, so they read the report S3 made instead of running S3 themselves.
s3_run(s3_example()).The first time, R sets up the matching Python library by itself (through reticulate), and the first run downloads the language model (420 MB). The R package and the Python library always have the same version, so they give the same results. Scopus API key: put SCOPUS_API_KEY=your-key in ~/.Renviron, restart R, then s3_enrich_citations(res).
Levels of analysis