BOOK Embeddings — Dialect Route
Multilingual-embedding service for the code-switched register of the Obsidian BOOK vault (Spanish address + Japanese romanji + Filipino martial vocab + English base — "romanji spanglish").
Serves a multilingual base model (paraphrase-multilingual-MiniLM-L12-v2),
deliberately NOT fine-tuned: fine-tuning on the mostly-English pair set was shown
to degrade dialect retrieval (7 -> 5 on the dialect slice), so this route keeps
the raw multilingual base, which scores best on the dialect bench.
Runs on Hugging Face ZeroGPU. Paired with the English-route Space
(book-embeddings, serving the tuned book-embedder-v1) behind ask_router.py.
Built by Jack of Shadows.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support