Getting a model

Any sentence-transformers-style ONNX export works: a directory that contains model.onnx and tokenizer.json. embeddings.hub/fetch-model downloads it from the Hugging Face Hub into ~/.cache/embeddings-clj. It tries onnx/model.onnx first, then model.onnx. It downloads a file again only when the cache does not hold it:

(require '[embeddings.hub :as hub])

(def model-dir (hub/fetch-model "sentence-transformers/all-MiniLM-L6-v2"))
;; opts: {:cache-dir "..." :revision "main" :variant ...}

Quantized exports: :variant :quantized tries the common quantized paths in this order: onnx/model_quantized.onnx, onnx/model_qint8_avx512_vnni.onnx, onnx/model_int8.onnx, and model_quantized.onnx. A string :variant names an explicit repo-relative .onnx path, for example "onnx/model_q4.onnx". Each variant is cached in its own subdirectory. Thus a full-precision copy and a quantized copy of the same model can both exist:

(hub/fetch-model "sentence-transformers/all-MiniLM-L6-v2" {:variant :quantized})

To fetch a model manually:

mkdir -p models/all-MiniLM-L6-v2
curl -fL -o models/all-MiniLM-L6-v2/model.onnx \
  https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2/resolve/main/onnx/model.onnx
curl -fL -o models/all-MiniLM-L6-v2/tokenizer.json \
  https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2/resolve/main/tokenizer.json

Compatible model families include all-MiniLM, all-mpnet, BGE, GTE, and E5. A model is compatible if its ONNX graph takes input_ids, attention_mask, and token_type_ids, and outputs token embeddings or a pre-pooled sentence embedding.

E5-style models expect instruction prefixes; pass :prefix per call:

(emb/embed model "how do I bake bread" {:prefix "query: "})
(emb/embed-batch model documents {:prefix "passage: "})