Skip to main content

EmbeddingProvider

clawft_kernel::embedding

Trait EmbeddingProvider 

Source 
pub trait EmbeddingProvider: Send + Sync {
    // Required methods
    fn embed<'life0, 'life1, 'async_trait>(
        &'life0 self,
        text: &'life1 str,
    ) -> Pin<Box<dyn Future<Output = Result<Vec<f32>, EmbeddingError>> + Send + 'async_trait>>
       where Self: 'async_trait,
             'life0: 'async_trait,
             'life1: 'async_trait;
    fn dimensions(&self) -> usize;
    fn model_name(&self) -> &str;

    // Provided methods
    fn embed_batch<'life0, 'life1, 'life2, 'async_trait>(
        &'life0 self,
        texts: &'life1 [&'life2 str],
    ) -> Pin<Box<dyn Future<Output = Result<Vec<Vec<f32>>, EmbeddingError>> + Send + 'async_trait>>
       where Self: 'async_trait,
             'life0: 'async_trait,
             'life1: 'async_trait,
             'life2: 'async_trait { ... }
    fn embed_query<'life0, 'life1, 'async_trait>(
        &'life0 self,
        query: &'life1 str,
    ) -> Pin<Box<dyn Future<Output = Result<Vec<f32>, EmbeddingError>> + Send + 'async_trait>>
       where Self: 'async_trait,
             'life0: 'async_trait,
             'life1: 'async_trait { ... }
    fn warm<'life0, 'async_trait>(
        &'life0 self,
    ) -> Pin<Box<dyn Future<Output = ()> + Send + 'async_trait>>
       where Self: 'async_trait,
             'life0: 'async_trait { ... }
}
Expand description

Trait for pluggable embedding backends.

Implementations convert text into fixed-dimensionality float vectors suitable for HNSW indexing and cosine similarity search.

Required Methods§

Source

fn embed<'life0, 'life1, 'async_trait>( &'life0 self, text: &'life1 str, ) -> Pin<Box<dyn Future<Output = Result<Vec<f32>, EmbeddingError>> + Send + 'async_trait>>
where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait,

Embed a single text chunk into a vector.

Source

fn dimensions(&self) -> usize

Dimensionality of the output vectors.

Source

fn model_name(&self) -> &str

Name of the embedding model (for metadata tracking).

Provided Methods§

Source

fn embed_batch<'life0, 'life1, 'life2, 'async_trait>( &'life0 self, texts: &'life1 [&'life2 str], ) -> Pin<Box<dyn Future<Output = Result<Vec<Vec<f32>>, EmbeddingError>> + Send + 'async_trait>>
where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait,

Embed a batch of text chunks.

The default implementation calls embed in a loop. Backends that support native batching should override this for efficiency.

Source

fn embed_query<'life0, 'life1, 'async_trait>( &'life0 self, query: &'life1 str, ) -> Pin<Box<dyn Future<Output = Result<Vec<f32>, EmbeddingError>> + Send + 'async_trait>>
where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait,

Embed a query (asymmetric retrieval lane).

Default = embed (symmetric backends: Mock, MiniLM, LLM). Providers with asymmetric training (e5 query: / passage:, Qwen3 instruction prefix) must override so graft/search queries land in the correct half of the space (WEFT-640 / ADR-059).

Source

fn warm<'life0, 'async_trait>( &'life0 self, ) -> Pin<Box<dyn Future<Output = ()> + Send + 'async_trait>>
where Self: 'async_trait, 'life0: 'async_trait,

Pre-warm the backend: run one throwaway embed so the model graph, runtime thread pool, and allocator are hot before the first real call. Best-effort — errors are swallowed (a backend with nothing to warm, like the Mock, just pays one cheap call). Intended to be called once at daemon startup so the first conversation turn doesn’t pay the model’s first-inference cost (ADR-058 Phase 5.3; mirrors the 6.2 STT pre-warm).

Dyn Compatibility§

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors§