Thyme 🌿

When did that happen, compared to everything else? Thyme lines up millions of dated events from open datasets on one scrolling timeline: album releases next to ball games, lives next to film premieres, books next to the wars they were written about. Pick an event and read every other event as years before or after it.

Built on TrailBase (SQLite with a REST API, a static file server and WebAssembly endpoints), a Python ingest, a few static pages hydrated with Alpine.js, and a small TypeScript app for the timeline itself.

What you get

• Timeline (/timeline.html): one horizontal band per data source, scrolled sideways across a world range of years, with a sticky year axis and a density minimap. Zoomed out, each band shows its top events per century, decade or year; zoomed in, every event. Filter by name, category, kind, source or domain; click an event for details and Compare from here to get a second axis of years before and after it. The view state is in the URL.
• Browse (/browse.html): a dozen domains (Film, TV & games, Music, Sports, Books, Awards, Science, Politics, Wars, Religion, Business, Arts, Exploration, Lives, Wikipedia lists), each a set of event categories and kinds with live counts; categories that are planned but empty show as stubs.
• Sources (/sources.html): every dataset with counts, year range, licence, last sync and what is planned.
• Search (/search.html): entities from every source plus the IMDB tables, each linking to its entity page and its place on the timeline.
• The original IMDB browser (movies, TV, people, genres, top rated) is still there, reached from the Film domain.

Quick start

# TrailBase v0.34+ (https://github.com/trailbaseio/trailbase/releases),
# Node 22.20+, Yarn 1.22, Python 3.11+, sqlite3
make check-deps
yarn install
trail run                 # once: creates traildepot/data and applies migrations; stop it again
make test-import          # small IMDB sample (or make import-data for the full datasets)
make sync-events          # every source under data/ (see Sources below)
make run                  # builds the WASM component and the site, serves on :4000

Visit http://localhost:4000. The pages call /api/records/v1/... and the custom endpoints on the same origin, so a separate static file server will not do. After adding data run make sync-events (and make rerank after a migration that changes the level-of-detail tables); the server can keep running, the ingest waits for it.

Sources

Every dataset is a source that contributes entities (a person, a team, a book) and events attached to them: instants (a release, a game, a medal) or spans (a life, a career, a run), each with a date precision and a certainty. Sources that are CSV files are TOML specs in scripts/ingest/specs/ (csvsource.py documents the format); anything needing logic is a Source subclass in scripts/ingest/sources/. Both register automatically.

Source	Files under data/	Events
imdb	derived from the imported IMDB tables	releases and series runs, episodes, cast and crew as participants, people's lives
musicbrainz	music/official_releases.csv	every official release (5M) on its artist's lane
wikidata_age	wiki/AgeDataset-V*.csv	1.2M lives with QIDs and occupation (the category)
lahman	sports/baseball/*.csv	MLB lives and careers, awards, All-Star games, Hall of Fame, franchises, World Series
nba	sports/nba/csv/*.csv	NBA lives, careers, drafts, every game since 1946, franchise eras
olympics	sports/olympic_events/athlete_events.csv	each Games, each athlete appearance, each medal
cricket	sports/cricket/*.csv	Tests, ODIs and T20Is since 1877
mls	sports/mls/matches.csv	every MLS match since 1996
bx_books	books/BX-Books.csv	271k books with year of publication, authors as participants
steam	steam/games_march2025_cleaned.csv	89k game releases, keyed by Steam app id
paintings	art/artist.csv	the lives of 421 famous painters
wikipedia_lists	wikipedia/lists/*.json	dated entries of Wikipedia bibliographies and lists (see below)
wikidata	wikidata/{items,labels}.jsonl.gz	people, wars, countries, awards from the dump (see below)

make sources        # every source, its row counts, last sync, whether files changed
make sync-events    # import the sources whose files changed
make link           # the same person across sources (also runs after every sync)
make rerank         # recompute rank and the level-of-detail tables

A sync fingerprints the source's files, skips unchanged sources, upserts rows by their stable per-source key, removes rows the source no longer produces and logs the run in source_syncs. Entities sharing an external identifier (IMDb, Baseball-Reference, MusicBrainz, NBA.com, Steam) or a Wikidata QID link with certainty; people whose name and years match link at lower confidence. Rank (votes, credits, how much else the owner did) decides what a zoomed-out view shows.

Wikipedia lists and bibliographies

make wikipedia-lists                                  # seeds: "Lists of books"
make wikipedia-lists SEEDS="Bibliography of Montana history" MAX_PAGES=1
make wikipedia-lists SEEDS="List of treaties" MAX_PAGES=20
make sync-events

The fetcher pulls the wikitext of bibliographies and "List of ..." pages through the MediaWiki API. The adapter turns {{cite book}}/{{cite journal}} templates and italic-title lines into books and articles with their authors, and every other dated line into a stub listed event, so a list of treaties or earthquakes is on the timeline before it has a parser of its own. Each list page is an entity that takes part in its entries.

Wikidata from the dump

pigz -dc latest-all.json.gz | python3 scripts/wikidata_extract.py -
python3 scripts/wikidata_extract.py latest-all.json.gz --limit 1000000    # a quick look
python3 scripts/wikidata_extract.py latest-all.json.gz --all-dated        # every dated item
make sync-events

The extract keeps humans, conflicts, countries and awards with their dates, identifiers and dated relations (awards received, positions held, conflict participation). With --all-dated any other item that has a calendar date (inception, publication, point in time) is kept as a dated stub under its first class, which is how every dated Wikipedia subject can be represented before it gets a parser.

Timeline API

The custom endpoints are a TrailBase WASM component built from wasm/src/ (make wasm): TypeScript against the trailbase-wasm SDK, bundled by Vite and componentized with jco. Editing a handler body needs a rebuild plus kill -HUP <trail pid>; adding a route needs a restart.

• GET /timeline?from&to&category&source&kind&span&q&participant&anchor&cursor&limit: one page of a window in (start, id) order, read by walking an index and stopping at limit; active are the notable spans still open at from; next is a keyset cursor; total is capped. Filters that the LOD tables cannot answer (a name, a kind) read the event stream, driven by the FTS hits when a name is given.
• GET /timeline/overview?from&to&bucket&per&category&source: the top per events per bucket (1, 10 or 100 years) across the requested categories, from the precomputed event_lod table, so a 500-year view costs what a decade costs.
• GET /timeline/density?from&to&bucket&by=source|category: counts per year bucket from event_density.
• GET /timeline/sources, GET /timeline/domains: what is loaded, with counts.
• GET /timeline/participants?event=, GET /timeline/links?entity=: who took part in an event; the same entity across sources.
• GET /search?q=&entities&titles&persons: entities from every source (trigram FTS, ranked by their top event) and the IMDB tables.

curl -g "http://localhost:4000/timeline?from=1960&to=1975&q=beatles&limit=50"
curl -g "http://localhost:4000/timeline/overview?from=1500&to=1999&bucket=10&per=8&category=astronomer,physicist"
curl -g "http://localhost:4000/timeline/density?from=-500&to=2030&bucket=10&by=source"

SQLite features in use

Feature	Where
STRICT tables, CHECK constraints, foreign keys with cascades	every table
FTS5 external-content indexes with unicode61 and highlight()	titles_fts, persons_fts, /search
FTS5 trigram tokenizer for substring name search	entities_fts, /timeline?q=, entity search
R*Tree (rtree_i32) over [start_year, end_year]	events_span, kept in sync by triggers
VIRTUAL generated columns (julianday)	events.start_julian, events.end_julian
JSON (json_each, json_object, json_valid)	genres, professions, events.detail
Window functions	/timeline, /timeline/overview, the LOD rebuild
Recursive CTE	span_lod (a span in every bucket it overlaps)
UPSERT, WITHOUT ROWID staging and LOD tables, partial indexes	ingest
ANALYZE, PRAGMA optimize, WAL	end of every sync

Join order is pinned with CROSS JOIN where the planner, after ANALYZE, would otherwise start from the seven-row sources table; source filters are checked on the events row, and a sparse source inside a dense window is paged on its own index first.

Development

make lint          # Biome, TypeScript (wasm/ and web/), ruff
make test          # Jest, the ingest and parser test suites
yarn build:site    # templates/ -> dist/, web/ -> dist/static/js/timeline.js

See CONTRIBUTING.md for the layout and conventions.

Licence

MIT for the code. Each dataset keeps its own licence, listed on the sources page.