gungorbasa/retrievalkit-minilm

RetrievalKit all-MiniLM-L6-v2 artifacts

This repository contains deterministic runtime exports used by RetrievalKit's optional embedding experiment. They are derived from sentence-transformers/all-MiniLM-L6-v2 at source revision c9745ed1d9f207416be6d2e6f8de32d1f16199bf.

Shared output contract

• BERT WordPiece tokenization
• maximum input length of 256 tokens
• masked mean pooling
• L2 normalization
• 384-dimensional F32 output
• cosine similarity

The ONNX models have dynamic batch and sequence axes. Callers must enforce the 256-token limit and pad batches only as needed.

Profiles

Profile	ONNX	Core ML
FP32	FP32 weights	FP32 ML Program
FP16	FP16 weights with F32 I/O	FP16 ML Program
Q8	dynamic signed-INT8 weights/activations	weight-only signed-INT8 with FP16 compute

All profiles return the same normalized F32 embedding contract, but different runtimes and precision profiles are not expected to be byte-identical. Applications should use the same profile for document and query embeddings unless a cross-profile combination has been separately qualified.

manifest-v1.json is the authoritative artifact inventory. It records the source revision, tokenizer files, input/output contract, sizes, quantization schemes, and SHA-256 digests. RetrievalKit clients pin an immutable repository revision rather than downloading from main.

These artifacts are experimental. No latency or retrieval-quality claim should be inferred from their presence; measured qualification is published separately in the RetrievalKit repository.

License

The source model declares Apache-2.0. RetrievalKit's conversion tooling and packaging are also Apache-2.0. See LICENSE, NOTICE, the upstream model card, and manifest-v1.json for provenance.