Crate laya_core 

Tensor-free core for Laya typed decisions.

Everything here is independent of the inference backend: prompt construction, question validation, routing signals, calibration and the question presets. The laya crate adds the candle model on top.

Keeping this crate free of tensor dependencies means a service that only needs routing or prompt construction — an API gateway deciding which checkpoint a request belongs to, say — can depend on it without pulling in candle or any GPU runtime.

use laya_core::lang;
use serde_json::json;

let detection = lang::analyse(&json!(
    "Mein Konto wurde zweimal belastet und ich möchte eine Rückerstattung"
));
assert_eq!(detection.language.as_deref(), Some("de"));
assert!(!detection.is_english);

The heuristic needs a few function words before it commits: a short German phrase such as "Mein Konto wurde zweimal belastet" is still reported as English, matching upstream exactly (its own Router docstring claims otherwise, but the implementation does not).

Re-exports

pub use error::Error;

pub use error::Result;

pub use question::Criteria;

pub use question::Question;

pub use question::QuestionKind;

pub use tokenizer::Tokenizer;

Modules

calibrate

Temperature calibration and the confidence metric, ported from upstream.

email

Email utilities for cleaning and structuring email inputs.

error

json_compat

json.dumps compatibility.

lang

Dependency-free script and language detection used to route between checkpoints.

presets

Ready-to-use question presets for common production decision workflows.

prompt

Prompt construction.

pycompat

Exact-parity helpers for the handful of places where Rust’s defaults differ from CPython’s. These are small, but every one of them is load-bearing for matching the upstream output byte for byte.

question

Typed question definitions, validated and rendered exactly as upstream does.

tokenizer

Checkpoint tokenizer.