Module embedding_e5 

e5-small-v2 ONNX embedding provider (WEFT-640).

Thin asymmetric wrapper over the BERT-family OnnxEmbeddingProvider: same WordPiece tokenizer, masked-mean pool, and L2 norm — the only behavioural difference is the e5 training convention of prepending passage: on documents and query: on queries.

• Dim: 384 (matches RVF/HNSW default and weave.toml target).
• License: MIT weights (intfloat/e5-small-v2); community ONNX exports.
• Staging: .weftos/models/e5-small-v2/model.onnx (+ sibling vocab.txt).
• CI: without the onnx-embeddings feature or staged artifacts the provider degrades to the ONNX hash fallback (deterministic, no download).

Selection: select_embedding_provider promotes e5 when its ONNX file is present and the runtime loads, below project-canonical Qwen3 (ADR-059) and above legacy MiniLM.

Structs

E5EmbeddingProvider

e5-small-v2 provider — 384-d BERT encoder with asymmetric query/passage prefixes.

Constants

E5_DIMS

Native output dimensionality (matches RVF / HNSW default).

E5_MAX_TOKENS

Max tokens for e5-small-v2 (BERT 512 context; leave room for specials).

E5_MODEL_NAME

Canonical model identifier (embedder_id / consistency contract).

PASSAGE_PREFIX

Document / passage prefix required by e5 training.

QUERY_PREFIX

Query prefix required by e5 training.