laya-decision

A pure-Rust port of Laya — a multilingual, non-autoregressive System-1 decision engine. Given a state (text, JSON, or a conversation list) and a set of typed questions (choice, score, noul), it scores every question in a single forward pass — no text generation. Inference runs natively on candle (CPU or Apple-Silicon Metal); routing, language detection, email cleaning, presets and shortlist are dependency-free.

The crate is published as laya-decision (the name laya was taken) but is imported as laya.

[dependencies]
laya-decision = "0.1"

use laya::{Router, RouteHints, triage_questions};
use serde_json::json;

let router = Router::with_defaults()?;
let state = json!({ "message": "I was charged twice, please refund" });
let result = router.predict(&state, &triage_questions(), &RouteHints::default())?;
println!("{}", serde_json::to_string_pretty(&result.to_json())?);
# Ok::<(), laya::LayaError>(())

The candle backend matches the upstream PyTorch model to within 1e-4 across all three checkpoints (English ModernBERT-large, multilingual mmBERT-base, typed-decisions). Enable the metal feature for Apple-Silicon GPU; disable default features for the weight-free pure-logic subset (routing, language detection, email cleaning).

See the repository for the HTTP server (laya-decision-serve), CLI (laya-decision-cli), and full documentation.

Apache-2.0. A port of upstream Laya (Apache-2.0); model checkpoints are distributed separately on the Hugging Face Hub under their own license.