Expose fastembed_model in config and a model dropdown on Admin → Models. Persist dimension from the chosen model, require restart to load it, and align legacy OpenAI default settings so fresh local-embedding installs start cleanly.
Per Stark