Crate remote_embedding_plugin 

Remote embedding provider for the graph-storage gear.

ADR-0004 names two real providers: the in-process ONNX default and “a second plugin [that] calls a remote inference endpoint”. This is the second one. It speaks the OpenAI-compatible POST /embeddings protocol, which is what OpenAI, Azure OpenAI, Groq, Together, Ollama, vLLM and most self-hosted inference servers expose, so one plugin covers the deployments that cannot run a model in the gear’s own process — a memory ceiling, a CPU budget, or a platform that already pays for an inference service.

What the identity can and cannot promise

The ONNX plugin names its embedding space by the SHA-256 of the bytes it loaded. A remote endpoint offers no bytes to hash: the only identity it has is which model, at which endpoint, at which width, and that is what this plugin declares. Two deployments pointing one model name at one host agree on the space; the same name at a different host, or a different requested width, do not. What no remote identity can catch is a vendor silently changing the weights behind a stable model name — ADR-0004 puts that under model governance rather than under the plugin, and it is the reason the ADR calls remote embedding governed data egress rather than an ordinary plugin call.

Vectors are normalized here

The gear’s index serves cosine similarity, and not every compatible endpoint returns unit vectors (OpenAI does, several self-hosted servers do not). Normalizing on this side makes the stored vectors comparable whatever the endpoint’s habit, and is part of the declared identity.

Structs

RemoteEmbeddingProvider

An OpenAI-compatible /embeddings endpoint, as the gear’s provider.

RemoteProviderConfig

What a deployment declares about its endpoint.

Enums

RemoteConfigError