Module backend 

Pluggable embedding backend abstraction.

Overview

The EmbeddingBackend trait decouples embedding generation from the transport layer (ONNX Runtime, Candle, static lookup, GGUF quantized). The engine selects a backend at startup via environment variables:

Env var	Backend chosen
DAKERA_BACKEND=onnx	OnnxBackend — production default (INT8 quantized)
DAKERA_BACKEND=candle	[CandleBackend] — pure-Rust Candle (requires candle feature)
DAKERA_BACKEND=static	StaticBackend — Model2Vec static lookup (~500× faster ingest)
DAKERA_BACKEND=gguf	[GgufBackend] — GGUF quantized via candle-gguf (requires candle feature)
(unset)	OnnxBackend — same as onnx

All backends implement the same EmbeddingBackend trait so callers are fully backend-agnostic.

Feature flags

• candle — enables [CandleBackend] and [GgufBackend] (requires HuggingFace Candle crates)

The onnx and static backends are always compiled — they have no additional dependencies beyond what the dakera-inference crate already requires.

Re-exports

pub use onnx::OnnxBackend;

pub use static_backend::StaticBackend;

Modules

onnx

ONNX Runtime embedding backend.

static_backend

Model2Vec static embedding backend.

Enums

BackendKind

Identifies which backend is active. Stored in HNSW node metadata so the background re-embed job knows which memories were indexed with a fast (static) embedding and need to be upgraded with the quality transformer.

Traits

EmbeddingBackend

Core embedding backend trait. All backends must be Send + Sync so they can be held behind an Arc and called from async tasks.

Functions

select_backend

Select an embedding backend based on environment variables.