Architecture

A Streamlit app, a SQLite file, and a handful of modules that each own one thing. No server of ours, no account, no background process.

radiowriter/
  __main__.py     the `radiowriter` command: starts the server, opens a browser
  app.py          the interface — the only module that imports streamlit
  paths.py        where the user's files live, per platform
  db.py           schema, additive migrations, every query
  pubmed.py       E-utilities: esearch, efetch, and the MEDLINE parser
  querybuilder.py composing a query from concept blocks
  issg.py         the four published ISSG search filters
  strategies.py   Radiopaedia heading → PubMed search strategy
  journals.py     reading the SCImago CSV
  semantic_scholar.py  citation counts
  unpaywall.py    is there a legal free copy, and where
  structure.py    the twenty-three article structures
  lint.py         Radiopaedia's linter rules
  radiopaedia.py  citations, numbering, Markdown → their HTML
  draft_io.py     a single draft in and out of a file
  backup.py       the archive in and out of a file
  data/           what was transcribed from Radiopaedia, with its date

app.py is the only place that knows about Streamlit. Everything below it is ordinary Python that can be called from a script and tested without a browser — which is what the check_*.py suite does.

The database

One SQLite file. Migrations are additive and never destructive: PRAGMA table_info tells us which columns are missing, ALTER TABLE ADD COLUMN adds them. The archive predates most of the features in it and has to keep working.

Table
articles	one row per PMID, with its state
screened_pmids	read and deleted, so they do not come back
lists, list_items	reading lists
drafts, draft_refs, draft_searches	the articles being written
searches	the history
journal_metrics, journal_issns	the SCImago file, as a table
citation_cache	a resolved citation is resolved forever
settings	email, keys, preferences

Journal metrics are a table rather than a lookup in memory because filtering by quartile has to be a JOIN. Doing it in Python would mean loading thirty thousand rows on every rerun and — worse — giving up SQL paging: the page count is computed on the filtered total, and only the database knows it.

Streamlit, and the two things it does that surprise people

A widget's state is discarded when the widget is not drawn. Hide a control behind a collapsed panel and its value silently reverts to the default. This is why the search filters are always on screen rather than inside an expander.

A widget's key cannot be written after the widget exists. Anything that changes a control's value from code has to happen in a callback, which runs before the widgets of that run are created — or the key has to change, which makes it a new widget that reads its value= again. Both idioms are used, and commented where they are.