src/memory/vector.rs

This file holds the two traits your semantic-retrieval memory is built from, plus dependency-free reference implementations behind the vector_index feature (0.3.1).

Since 0.4.0 the substrate has real tiers behind it: OllamaEmbedder and OpenAiEmbedder (embeddings), VectorMemoryStore (vector_memory), the loopctl-hnsw companion, and the loopctl-vector external backends.

Key items

EmbeddingProvider takes text in and gives you an Embedding, which is a vector plus its dimension.

Required on you: dim() and embed(&str).

Its provided method embed_batch loops embed and preserves your input order.

VectorIndex is a nearest-neighbour store keyed by Uuid. add upserts, and search returns the cosine-scored top-k.

Results sort descending with an id tiebreak, so your order is deterministic, and remove is idempotent. You also get len() and is_empty().

A dimension mismatch surfaces as LoopError::Memory.

LinearVectorIndex is a brute-force O(n) cosine scan with RwLock interior mutability. It is deliberately simple: it is the correctness oracle any faster index must reproduce.

HashingEmbedder is a deterministic hash-based embedder. It uses no network and no key, and the same text yields the same vector.

Its retrieval quality is unsuitable for real use, which makes it exactly right for your tests. Zero-token input yields the documented zero vector.

cosine_similarity is the free helper every implementation shares.

Behavior notes

Both traits are object-safe async, with boxed futures exactly like LoopMemory.

So a store can hold Box<dyn EmbeddingProvider> and Box<dyn VectorIndex> and swap backends without touching its contract.

You can run this end-to-end without any provider: the vector-index-demo example and the integration suite exercise the whole layer on the reference implementations.

Deep dive: Memory, the vector-primitives section.