Building the dataset

The dataset is compiled by the builder CLI from two committed layers, so a rebuild can never clobber authored work:

Layer	Who writes it	Holds
`config/overrides/`	you (hand-edited)	Structure the open sources can’t express: Series / Franchise boundaries, ordering, `alternateCutOf`, `watchOrders`, which series are linearly `numbered`.
`data/`	`builder build` (generated)	The resolved records: overrides + facts from open data + computed `absoluteNumber`. Never hand-edit.

(absoluteNumber is the running episode count across a whole series, computed by the builder for linearly-numbered series.)

Because builder build treats overrides/ as read-only input, builds are deterministic and idempotent: the same overrides plus the same pinned sources always produce the same data/.

Prerequisites

Go (see the module’s go.mod for the version). Build the CLI from the repository root:

go build ./cmd/builder

Commands

./builder init                 # download the pinned sources into .sources/
./builder build                # (re)build data/ for all overrides
./builder build demon-slayer   # build/rebuild just one franchise or series
./builder refresh              # update sources to latest, bump pins, rebuild all

The build fails on any unknown id, dangling reference or schema violation, so a successful build is always a valid dataset. Where it makes a low-confidence guess (chiefly title-language tagging) it prints a report — pin those cases with an override.

Adding an entry

Create config/overrides/series/<id>.yaml describing the structure, then run ./builder build <id>. The builder fills in facts (titles, episode counts, cross-IDs) from the open sources and writes the resolved record to data/series/<id>.yaml.

Sources

Facts come from open data sources, each on its own terms. AniList is not used: its ToS forbids storing or redistributing its content.

anime-offline-database (ODbL + DbCL) — titles, season / year, episode counts, cross-IDs.
Anime-Lists/anime-lists (no stated licence; bare ID pairs) — AniDB↔TVDB mapping and movie-set grouping.
Wikidata (CC0) — character & staff names, resolved by QID.

What each source contributes, and what the terms mean for anyone redistributing the result, is set out in Sources and licensing.

Sources are not committed. builder init downloads them into a gitignored .sources/ cache at the versions pinned in config.yaml, and builder refresh updates those pins. (The build verifies pinned checksums, so a rebuild uses exactly the sources it was authored against.)

Serving it

The same dataset is compiled into the API binary with go:embed, so the server is stateless and self-contained:

go run ./cmd/api                 # listens on :8080

See Using the API for the request format.

Using the dataset Coverage