Take it with you
The corpus is public data, not only a public endpoint. Every row — the vector, the pipeline identity, the recording claims, the contributor count — is dedicated to the public domain under CC0 1.0, and published weekly as files anyone can download, load into their own instance, or build on. The reasoning is ADR-0013.
Latest export · 2026-09-15
25,886 embeddings across
1 pipeline identity,
and 23,198 recording claims. Generated
2026-09-15T23:03:20Z. A new one lands every Sunday; if this date is more
than a fortnight old, the timer has stopped and the page is telling you so.
| File | Pipeline identity | Rows |
|---|---|---|
embeddings-laion_clap_htsat_unfused_frontend1_artifact1_pool1_fp32.csv.gz |
laion/clap-htsat-unfused+frontend1+artifact1+pool1+fp32 |
25,886 |
claims.csv.gz |
recording claims — fingerprint_hash, recording_mbid, claim_count |
23,198 |
manifest.json |
this listing, as data — /export/latest.json always redirects to the current one |
What is in a file
One gzipped CSV per pipeline identity, so you take only the vectors you can compare with.
Columns: fingerprint_hash (the SHA256 of an AcoustID fingerprint — the key),
named (whether any client has claimed a recording id for it),
pipeline_version, contributor_count (how many independent
installations confirmed the vector), created (the day it arrived), and
embedding (512 floats, as pgvector prints them). The claims file maps hashes
to MusicBrainz recording ids with a count per pair.
What is never in a file: which installation contributed a row. The
agreement count is published; the parties are not. A random client_id
identifies nobody, but the set of recordings one contributed is a person's record
collection, and that stays on this server (ADR-0013 point 4).
Run your own
The server is MIT and runs from the repository's compose files; an export loads into a fresh instance and it answers the same lookups under the same keys. A mirror is another instance, not a federation — nothing syncs, and a copy is as current as the export it loaded. The runbook has the steps.
What a snapshot cannot do
A row deleted here — a takedown, a retraction — is absent from the next export, and a takedown regenerates the export straight away rather than at the next tick. What nobody can do is recall a file already downloaded: that is what public domain means, and it is stated here rather than discovered later (ADR-0013 point 6). Exports older than five weeks expire; the first of each month is kept.