Crate sie_sdk 

Rust client for the SIE inference server.

use sie_sdk::{Client, Item, OutputType};

let client = Client::new("http://localhost:8080")?;

let result = client
    .encode("BAAI/bge-m3", [Item::text("Hello world")])
    .output_types([OutputType::Dense, OutputType::Sparse])
    .send_one()
    .await?;

println!("{:?}", result.dense);

Waiting for capacity

SIE scales from zero and loads models on demand, so a request can legitimately be answered with “not yet”. The client absorbs that: PROVISIONING, MODEL_LOADING, LORA_LOADING and RESOURCE_EXHAUSTED are retried inside a wall-clock budget (15 minutes by default) with bounded, jittered backoff. Endpoints that are not idempotent never replay a request that may already have reached a worker.

Opt out per call when a failure is more useful than a wait:

let result = client
    .encode("BAAI/bge-m3", [Item::text("Hello")])
    .wait_for_capacity(false)
    .max_oom_retries(0)
    .send_one()
    .await;

if let Err(error) = &result
    && error.is_capacity_error()
{
    // Fall back to a smaller model, or shed the request.
}

Re-exports

pub use client::WatchMode;

pub use client::jobs::connection_name;

pub use client::jobs::require_connection_name;

pub use client::jobs::require_connector_idempotency_key;

pub use client::ChunkStream;

pub use client::Client;

pub use client::ClientBuilder;

pub use error::Error;

pub use error::ModelLoadErrorClass;

pub use error::Result;

pub use error::TransportErrorKind;

pub use media::Samples;

pub use retry::RequestOptions;

pub use scoring::maxsim;

pub use scoring::maxsim_batch;

pub use types::AudioInput;

pub use types::BinaryInput;

pub use types::CapacityInfo;

pub use types::Classification;

pub use types::DType;

pub use types::DetectedObject;

pub use types::EncodeResult;

pub use types::Entity;

pub use types::ExtractResult;

pub use types::HealthResponse;

pub use types::ImageInput;

pub use types::Item;

pub use types::ModelInfo;

pub use types::ModelState;

pub use types::Multivector;

pub use types::OutputDType;

pub use types::OutputType;

pub use types::RequestMetadata;

pub use types::RequestUsage;

pub use types::ScoreResult;

pub use types::SparseVector;

pub use types::TimingInfo;

Modules

blocking

A blocking facade over the async client.

client

The client, its builder, and the one request path every endpoint shares.

error

The single error type returned by every fallible SDK operation.

media

Turning caller-supplied images, audio and documents into the bytes the wire carries.

ndarray

Conversions into ndarray types, for callers whose numeric code already uses it.

redaction

Scrubbing secrets out of anything that reaches a log.

retry

The retry state machine shared by every endpoint.

scoring

Client-side MaxSim scoring for late-interaction (ColBERT-style) models.

types

Wire types for every SIE endpoint.

wire

Server error envelopes: decoding them, and mapping status + code onto Error.

Constants

VERSION

The version this SDK reports to the server.