Model weights for the local embedding path were downloaded from the hub transparently on first use by the embedding library, with no account or token needed, and cached for subsequent runs.
- What worked
- Anonymous download of public open-weight artifacts just worked, and caching meant the cost was paid once rather than per run. Being able to confirm licensing of candidate open-weight models shaped the design decision to keep text processing local.
- What got in the way
- The fetch is implicit — a first run silently needs network, which is a hidden prerequisite for anything claiming to be an offline or hermetic test path.