skosdex

Use when (re)building skosdex concept embeddings for many or all schemes at once — a one-off GPU batch box to embed the whole corpus ASAP, or a parallel CI matrix for incremental per-scheme additions. Covers the embed pipeline, where to run it (and why not the prod box), model parity, committing + deploying the vectors with no blackout, and — critically — tearing the GPU box down afterwards so it stops billing.

Use when working with skosdex concept embeddings — the per-scheme and combined vectors, their JSON formats, the in-browser "similar by meaning" / viz consumers, or the server-side similarity REST API (tools/embed_api.mjs). Covers how to build, combine, query, and extend them.

Use when redeploying the live skosdex stack (skosdex.fly.dev — Oxigraph + Solr + web) on Fly.io, choosing between a direct deploy and the zero-downtime cutover, forcing/monitoring a deploy, or recovering a broken one. The deploy is CI-driven from claude/main.

Use when adding a new SKOS concept scheme / vocabulary / thesaurus to the skosdex repository (creating its third_party/skos/<slug>/ folder and meta.ttl), or when deciding whether a scheme's data may be committed and bundled given its license.

Use when discovering, researching, triaging, or maintaining the list of candidate SKOS vocabularies for skosdex — i.e. updating third_party/skos/CANDIDATES.md, verifying a scheme's license, or deciding whether a vocabulary can be bundled or must stay metadata-only.

Use when producing the shippable skosdex dataset — merging openly-licensed canonical N-Quads into dist/bundle.nq.gz, generating dist/solr-docs.json, building the docs/ site, or understanding which schemes are included/excluded and why.

Use when fetching, parsing, normalizing, or canonicalizing a SKOS scheme into N-Quads in skosdex — i.e. running the scripts/skosdex fetch/normalize/canonicalize pipeline or debugging its normalized.nq.gz / canonical.nq.gz output, including the gzip + Git LFS storage and zip/gz source decompression.

Use when bringing up, configuring, or troubleshooting the skosdex services — the Oxigraph SPARQL endpoint, the Solr index, and the web frontend via docker compose (or swapping Oxigraph for Apache Jena Fuseki).