Crate atomr_agents_embed

Expand description

EmbeddingActor + pluggable AnnIndex.

Structs§

EmbeddingToolStrategy: Embeds tool descriptors at construction; at select-time, embeds the user turn and returns top-k tools that fit the budget.
Hit: A single vector-store search result.
InMemoryAnnIndex: Linear-scan, cosine-similarity index. Mirrors the algorithm in atomr_accel_agents::CpuVectorIndex but lives in-process to avoid requiring an ActorSystem for unit tests.
InMemoryVectorStore: In-memory cosine-similarity vector store. Applies MetadataFilter at query time before ranking. Suitable for tests and small hot tiers.
MockEmbedder: Deterministic, dimension-configurable mock embedder. Produces stable vectors so tests are reproducible.

MetadataFilter: A set of equality / range predicates over metadata keys, honored by the backend at query time. And composes sub-predicates conjunctively. MetadataFilter::matches is provided for in-memory backends and as a documented fallback for backends that cannot push the predicate down to the engine.

AnnIndex: Approximate-nearest-neighbor index. v0 ships an in-memory linear scan; a CUDA / external-service implementation slots in behind the same trait without changing call sites.
Embedder: Single text → vector. The default implementation in production will wrap an atomr-infer ModelRunner configured with an embedding model; for unit tests, MockEmbedder produces deterministic vectors from a hash.
Embeddings: Batch embeddings interface. A blanket impl bridges any existing single-text Embedder, so all current embedders satisfy Embeddings for free.
VectorStore: Backend-agnostic vector store: dense upsert, filtered top-k query, and delete. In-memory impl lives in atomr-agents-embed; pgvector and Redis backends live in this crate behind feature flags.