oonnx2oracle
Reference R3

Model matrix

The full set of registered presets. Click any column header to sort — useful when you're picking by size, dimension, or pooling strategy. Refresh real-DB evidence with scripts/check_model_compatibility.py --all-presets.

Preset Task HuggingFace repo Dimensions Size (FP32) Pooling Oracle name
all-MiniLM-L6-v2 embedding sentence-transformers/all-MiniLM-L6-v2 384 ~90 MB mean ALL_MINILM_L6_V2
all-MiniLM-L12-v2 embedding sentence-transformers/all-MiniLM-L12-v2 384 ~130 MB mean ALL_MINILM_L12_V2
all-mpnet-base-v2 embedding sentence-transformers/all-mpnet-base-v2 768 ~420 MB mean ALL_MPNET_BASE_V2
bge-small-en-v1.5 embedding BAAI/bge-small-en-v1.5 384 ~130 MB cls BGE_SMALL_EN_V1_5
nomic-embed-text-v1 embedding nomic-ai/nomic-embed-text-v1 768 ~540 MB mean NOMIC_EMBED_TEXT_V1
ms-marco-MiniLM-L-6-v2 reranker cross-encoder/ms-marco-MiniLM-L-6-v2 ~90 MB MS_MARCO_MINILM_L_6_V2
ms-marco-MiniLM-L-12-v2 reranker cross-encoder/ms-marco-MiniLM-L-12-v2 ~130 MB MS_MARCO_MINILM_L_12_V2

Reranker presets

The two ms-marco-MiniLM-* entries above are cross-encoder rerankers, not embedding models. Oracle registers them with function:"regression" metadata; you query them via PREDICTION(model USING q AS DATA1, d AS DATA2) rather than VECTOR_EMBEDDING. Output is a scalar logit — higher means more relevant. Apply 1 / (1 + EXP(-score)) if you need a [0,1] probability.

SentencePiece-based rerankers like BAAI/bge-reranker-base are not supported (same constraint as the embedding path — Oracle's BertTokenizer op can't represent SentencePiece vocabularies). The cross-encoder/ms-marco-MiniLM-* family uses WordPiece and works first-class.

A note on sizes

The FP32 size is the on-disk footprint of the augmented ONNX — encoder plus tokenizer ops plus pooling plus L2 norm. The encoder alone is about 85% of that. These are the models as downloaded, before any Oracle-side compression or tablespace overhead.

Dimensions and storage

The output dimension drives disk usage for the vector column:

Oracle's HNSW index doubles that rough figure once built. Cosine distance and dot product rank identically on L2-normalized vectors — all listed presets normalize, so use whichever VECTOR_DISTANCE metric reads cleaner in your SQL.