isidorus-ai/multilingual-e5-small-coreml

Isidorus embedder (multilingual-e5-small) — Core ML

FP16 mlprogram Core ML conversion of intfloat/multilingual-e5-small for the Isidorus app — registry label embedding_coreml, tier: Apple platforms (embedding_coreml).

• License: MIT (upstream model license)
• Source checkpoint: intfloat/multilingual-e5-small @ 614241f622f53c4eeff9890bdc4f31cfecc418b3
• Artifact: embedding.mlpackage.zip — the uncompiled distribution format; compile once on device with MLModel.compileModel(at:) and persist the resulting .mlmodelc (a local build artifact, never redistributed).
• Deployment floor: iOS 16 / macOS 13 (mlprogram).

Model interface

Tensor	Type	Shape	Notes
input_ids	Int32	[1, 512]	truncated & padded to 512
attention_mask	Int32	[1, 512]	0/1
token_type_ids	Int32	[1, 512]	all zeros (single-sentence; BertModel segment ids)
embedding (out)	Float16	[1, 384]	L2-normalized CLS vector

Tokenization stays in the app (Rust, HuggingFace tokenizers crate) against the bundled tokenizer.json. Pooling is CLS (last_hidden_state[:, 0, :]) + L2 normalization, matching the ONNX embedder.

On-device usage (Swift)

let pkgURL  = extractedModelDir.appendingPathComponent("embedding.mlpackage")
let compiled = try await MLModel.compileModel(at: pkgURL)   // one-time
let dest = appSupport.appendingPathComponent("models/embedding_coreml/embedding.mlmodelc")
try? FileManager.default.removeItem(at: dest)
try FileManager.default.moveItem(at: compiled, to: dest)    // persist at a stable path

let config = MLModelConfiguration()
config.computeUnits = .all
let model = try MLModel(contentsOf: dest, configuration: config)

Reproducibility

See manifest.json (source revision, tool versions, shapes, SHA-256) and parity_report.json (torch FP32 vs Core ML FP16 embeddings; min cosine 0.999960, max abs diff 0.0016).

Converted with tools/convert_embedder_coreml.py in the Isidorus workspace.