Expand description
Local CPU sentence-embedding Embedder for
gonzalo, built on Candle + all-MiniLM-L6-v2 (ADR 0013).
CandleEmbedder::load resolves the model weights (via an
EmbedderConfig::model_path override, else a one-time anonymous hf-hub
download) and loads the BERT model + tokenizer once on CPU. embed then
tokenizes, runs the forward pass, masked-mean-pools the token states, and
L2-normalizes to a 384-dim unit vector. The synchronous CPU forward runs
inside spawn_blocking so it never blocks the async runtime.
Structsยง
- Candle
Embedder - A local CPU sentence embedder (Candle + all-MiniLM). Cheap to clone.
- Embedder
Config - Configuration for
CandleEmbedder::load.