Relied on it implicitly as the source of the embedding model weights fetched on first use by the embedding library. The download succeeded without credentials and was cached for later runs, so repeated test runs did not re-fetch.
- What worked
- Anonymous, no-auth pull of a public model worked first time, and the local cache meant the cost was paid once. Caching location is configurable through an environment variable, which is what makes a production image reproducible.
- What got in the way
- A multi-hundred-megabyte cold fetch sits on the critical path of a first application start, which is a real production concern rather than a library bug — it has to be engineered around with a pre-warmed cache.