Crate uni_xervo 

Unified Rust runtime for local and remote embedding, reranking, and generation models.

Uni-Xervo provides a single, provider-agnostic API for loading and running ML models across a wide range of backends — from local inference engines (Candle, ONNX Runtime, mistral.rs) to remote API services (OpenAI, Gemini, Anthropic, Cohere, Mistral, Voyage AI, Vertex AI, Azure OpenAI), and raw ONNX graphs.

Key concepts

• ModelRuntime — the central runtime that owns providers and manages a catalog of model aliases.
• ModelAliasSpec — a declarative specification that maps a human-readable alias (e.g. "embed/default") to a concrete provider + model pair.
• Providers — pluggable backends that implement ModelProvider. Each provider advertises the tasks it supports and knows how to load models.
• Traits — EmbeddingModel, RerankerModel, and GeneratorModel are the task-specific interfaces returned by the runtime. The retrieval surface adds SparseEmbeddingModel, MultiVectorEmbeddingModel, and HybridEmbeddingModel (single forward pass over a multi-output graph), and the multimodal extension surface adds ImageEmbeddingModel, AudioEmbeddingModel, MultimodalEmbeddingModel, NlpModel, DocumentExtractionModel, TranscriptionModel, and OcrModel — resolved via the matching methods on ModelRuntime (sparse_embedder, multi_vector_embedder, hybrid_embedder, image_embedder, multimodal_embedder, nlp_model, document_extractor, transcriber, ocr_model). local/onnx implements the sparse / multi-vector / hybrid / image / NLP / OCR tasks and remote/cohere + remote/gemini implement multimodal embedding; audio_embedder has no bundled provider yet.

Quick start

use uni_xervo::api::{ModelAliasSpec, ModelTask};
use uni_xervo::runtime::ModelRuntime;
use uni_xervo::provider::candle::LocalCandleProvider;

let spec = ModelAliasSpec {
    alias: "embed/local".into(),
    task: ModelTask::Embed,
    provider_id: "local/candle".into(),
    model_id: "sentence-transformers/all-MiniLM-L6-v2".into(),
    revision: None,
    warmup: Default::default(),
    required: true,
    timeout: None,
    load_timeout: None,
    retry: None,
    options: serde_json::Value::Null,
};

let runtime = ModelRuntime::builder()
    .register_provider(LocalCandleProvider::new())
    .catalog(vec![spec])
    .build()
    .await?;

let model = runtime.embedding("embed/local").await?;
let embeddings = model.embed(&["Hello, world!"]).await?;
println!("dim: {}", embeddings.vectors[0].len());

Modules

api

Public API types for configuring models, catalogs, and runtime behavior.

cache

Model and weight cache directory resolution.

error

Error types for the Uni-Xervo runtime.

prelude

Convenience re-exports for the common uni-xervo surface.

provider

Provider implementations for local and remote model backends.

reliability

Reliability primitives: circuit breaker, instrumented model wrappers with timeout and retry support, and metrics emission.

runtime

The core runtime that manages providers, catalogs, and loaded model instances.

score

Host-side similarity scoring for sparse and multi-vector embeddings.

traits

Core traits that every provider and model implementation must satisfy.