src/provider/embeddings.rs

This file holds the two EmbeddingProvider backends a real semantic-memory consumer needs, one local and one frontier.

Each sits behind the provider feature it belongs to plus vector_index, so there are no new features and default = [] is untouched.

Key items

OllamaEmbedder talks to Ollama’s native /api/embed endpoint, and a chat-profile /v1 suffix is normalized away.

It defaults to nomic-embed-text at 768 dims and is key-free. from_env() requires OLLAMA_EMBEDDING_MODEL and deliberately refuses OLLAMA_MODEL, and OLLAMA_EMBEDDING_DIM optionally pins the geometry for you.

OpenAiEmbedder calls /v1/embeddings and defaults to text-embedding-3-small at 1536 dims.

with_dimensions(256..=3072) rides the 3-series Matryoshka truncation only when you set it, and the key is redacted in Debug.

Both give you validated builders with the family’s shared HTTP config and embed_texts batching at 256 inputs or about 250k estimated tokens per request.

Both read responses with a bound derived from your dimension, offer optional normalized() for exact L2 in place, and emit per-batch loopctl.embed.batch metric events.

Behavior notes

Dimension honesty is enforced. A server answer whose length differs from your configured dim() fails as a config-validation error naming both numbers, because embedding models are not interchangeable across a store’s lifetime.

OpenAI responses are reassembled by data[].index and never by array position, and duplicate or out-of-range sets refuse.

Every input guards against the 8191-token ceiling before the wire, naming the input index to you.

Component finiteness is enforced on every answer, normalized or not, and non-finite answers fail the embed.

For the deep dive, see Embeddings.