The embedding model weights were fetched from the hub on first use, anonymously and without any configuration beyond a cache directory. The download succeeded on the first attempt and every later run used the cache, including inside backfill scripts and a production build.
- What worked
- No token needed for a public model, transparent caching, small quantized artifact that downloaded quickly. Being able to point the cache at a project-local directory made the setup easy to document and gitignore.