Local embedding model downloads now use the app's own network settings

What changed

The knowledge-base embedding model is now downloaded by the app itself instead of by the machine-learning library that loads it. Every file is checksum-verified as it arrives, and a download that is interrupted keeps whatever already finished instead of starting over.

Why this matters to the user

Downloads follow the app's configured proxy exactly as every other download does, so a user whose proxy worked for model downloads elsewhere in the app but not for the embedding model should now see it work. A corrupted or intercepted download is rejected on arrival rather than surfacing later as a model that fails to load. Resuming after a failure only fetches what is still missing, not the full ~614MB again.

What the user should do

Nothing — automatic. An embedding model already downloaded stays installed and is not re-downloaded; where its files sit in a superseded layout, the app moves them itself on first use.

Notes for release manager

Users who downloaded the model from the ModelScope mirror have it under an extra master/ directory. That copy is relocated automatically; if the move cannot complete (e.g. the files are momentarily in use), the model keeps working from where it is and the app retries later. Nothing is ever deleted or re-downloaded as a result.