Expand description
Embedding storage codec + similarity. Always compiled (the library indexes and searches vectors even when built without the HTTP LLM backend).
Vectors are stored in the container’s embeddings/vectors.f32 asset as a
flat, row-major little-endian f32 matrix of shape rows × dim.
Functions§
- centroid
- Mean of a set of vectors (centroid). Empty input → empty vector.
- cosine
- Cosine similarity in [-1, 1]; 0 if either vector is degenerate or mismatched.
- decode_
f32_ matrix - Decode a flat little-endian f32 buffer into
rows × dim. - encode_
f32_ matrix - Encode an embedding matrix (
rows × dim, row-major) to little-endian f32 bytes.