Feature request: Support multilingual embedding models (non-English sessions)

Thanks a lot for open-sourcing funes — the "raw evidence with provenance" design is exactly what's been missing in this space.

I noticed the default embedding model is pinned to BAAI/bge-small-en-v1.5, while my sessions are mostly in Chinese, so retrieval quality is a concern for me.

I understand that pinning the embedding model ensures a consistent vector space across machines. Given that design, is there any plan to let users choose (or override) the embedding model locally — for example, via a flag like --embedding-model, with the model identity stamped into the memory as it is now?

Alternatively, would you consider supporting a multilingual model (e.g., bge-m3 or a multilingual-e5 variant) as the default or as an official alternative? Happy to hear your thoughts.

Start by locating the current default BAAI/bge-small-en-v1.5 configuration and where the model identity is stamped into memory. Clarify whether the intended outcome is a --embedding-model override, an official multilingual alternative, or both, then define acceptance criteria for model selection and persisted identity.

Written by the indexing model from the issue text.