Struct LocalProvider 

pub struct LocalProvider(/* private fields */);

Available on crate feature semantic only.

Local ONNX embedding provider.

Cheap to clone (Arc under the hood). All inference is serialized behind an internal mutex so the same provider can be shared across threads — for high-throughput scenarios, use one LocalProvider per worker.

Implementations

impl LocalProvider

pub fn builder() -> LocalProviderBuilder

Start a LocalProviderBuilder.

Use the builder when you need a non-default pooling strategy, custom prefixes, custom max sequence length, or a hand-tuned thread count.

pub fn from_pretrained(model_id: &str) -> Result<Self>

Construct from a Hugging Face Hub model id.

Downloads the ONNX model + tokenizer on first use via hf-hub, caching to the user’s HF cache directory. Looks up the appropriate pooling strategy and query / document prefixes from internal tables (BGE → Cls + query prefix, E5 → Mean + query/ passage prefixes, …).

Arguments

• model_id — HF identifier such as "BAAI/bge-small-en-v1.5".

Errors

Returns crate::Error::Config if the repo lacks an ONNX file or tokenizer.json; crate::Error::Onnx for ort load failures; crate::Error::Tokenizer for tokenizer parse failures.

Example

use txtfp::semantic::LocalProvider;
let p = LocalProvider::from_pretrained("BAAI/bge-small-en-v1.5")?;

pub fn from_onnx( onnx_path: &Path, tokenizer_path: &Path, pooling: Pooling, ) -> Result<Self>

Construct from explicit ONNX + tokenizer paths.

Use this for self-hosted models or air-gapped deployments where downloading from the HF Hub is not an option.

Arguments

• onnx_path — path to a .onnx graph file.
• tokenizer_path — path to a tokenizer.json (HF tokenizer format).
• pooling — output pooling strategy (BGE → Pooling::Cls, E5/MiniLM → Pooling::Mean).

Errors

Returns crate::Error::Onnx if the model fails to load, crate::Error::Tokenizer if the tokenizer JSON is invalid.

Example

use std::path::Path;
use txtfp::semantic::{LocalProvider, Pooling};

let p = LocalProvider::from_onnx(
    Path::new("/srv/models/embedder.onnx"),
    Path::new("/srv/models/tokenizer.json"),
    Pooling::Cls,
)?;

pub fn embed_document(&self, input: &str) -> Result<Embedding>

Embed input as a document.

Prepends the model’s document prefix (e.g. "passage: " for E5) when applicable, then runs tokenize + inference + pooling. The returned Embedding carries model_id = Some(...) so downstream comparisons can detect cross-model leaks.

Errors

See LocalProvider::embed for the error variants.

pub fn embed_query(&self, input: &str) -> Result<Embedding>

Embed input as a query.

Prepends the model’s query prefix (e.g. "Represent this sentence for searching relevant passages: " for bge-*, "query: " for e5-*). For models that don’t use asymmetric encoding, this is identical to embed_document.

Use embed_query for the search side of a retrieval pipeline and embed_document for the corpus side.

Trait Implementations

impl Clone for LocalProvider

fn clone(&self) -> LocalProvider

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for LocalProvider

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl EmbeddingProvider for LocalProvider

type Input = str

The kind of input this provider consumes.

fn embed(&self, input: &str) -> Result<Embedding>

Compute an embedding for input.

fn model_id(&self) -> &str

The model identifier this provider produces.

fn dimension(&self) -> usize

The output dimensionality.