com.sky.sentis.e5

sentis-e5

Multilingual E5 small text embedder converted to Unity Inference Engine (Sentis) FP16.

Files

e5_small_fp16.sentis   # encoder (384-dim output)
tokenizer.json         # HuggingFace tokenizer
E5Embedder.cs          # self-contained Unity Sentis inference

Inference

Prefix queries with query: and passages with passage: . The sentence embedding is the attention-masked mean of last_hidden_state, L2-normalized (so relevance is a plain dot product).

A complete self-contained implementation lives in E5Embedder.cs (tokenization, masked mean-pool, L2-normalize, sync + async readback). Minimal usage:

var embedder = new E5Embedder(BackendType.CPU);
embedder.Load(modelRoot); // folder holding e5_small_fp16.sentis + tokenizer.json
float[] v = embedder.EmbedText("how tall is the eiffel tower", isQuery: true); // 384-d, L2-normalized

For a GPU backend use EmbedTextAsync; a synchronous GPU readback can stall or crash on some mobile GPUs.

License & attribution

MIT. Converted from intfloat/multilingual-e5-small (MIT).