Struct SemanticModel 

pub struct SemanticModel { /* private fields */ }

Wrapper around an embedding model backend.

Supports multiple backends selected at compile time via feature flags. When multiple backends are available, ORT is preferred (higher quality); model2vec is used as a fallback.

Implementations

impl SemanticModel

pub fn load() -> Result<Self>

Load the model from the OS cache directory.

Tries backends in priority order: ORT (highest quality), then model2vec (no ONNX dependency, Windows-friendly).

Errors

Returns an error if no backend is available or loading fails.

pub fn model_cache_path() -> Result<PathBuf>

Return the OS-appropriate cache path for the ORT model file.

Uses dirs::cache_dir() / bones / models / minilm-l6-v2-int8.onnx.

Errors

Returns an error if the OS cache directory cannot be determined.

pub fn model_cache_root() -> Result<PathBuf>

Return the OS-appropriate cache root for model files.

Errors

Returns an error if the OS cache directory cannot be determined.

pub fn is_cached_valid(path: &Path) -> bool

Check if cached model matches expected SHA256.

pub fn extract_to_cache(path: &Path) -> Result<()>

Extract bundled model bytes to cache directory.

Errors

Returns an error if no bundled bytes are available, or if the file system operations fail, or if the extracted model fails SHA256 verification.

pub fn dimensions(&self) -> usize

The dimensionality of embedding vectors this model produces.

pub fn backend_id(&self) -> &'static str

A stable identifier for the active backend, used to detect backend switches that require re-embedding stored vectors.

pub fn embed(&self, text: &str) -> Result<Vec<f32>>

Run inference for a single text input.

Errors

Returns an error if the runtime is unavailable or inference fails.

pub fn embed_batch(&self, texts: &[&str]) -> Result<Vec<Vec<f32>>>

Batch inference for efficiency.

Errors

Returns an error if the runtime is unavailable or batch inference fails.