Crate laya 

Laya typed decisions in Rust.

An independent Rust port of the Laya decision models: a bidirectional encoder answers constrained questions — a choice, a rubric score, or the probability of a proposition — in one forward pass, with no token-by-token decoding and no generated JSON.

use laya::Agent;
use serde_json::json;

let agent = Agent::from_pretrained("aac6fef/laya-mlx")?;
let result = agent.predict(
    &json!("I was billed twice. Please refund the duplicate."),
    &json!({
        "department": {
            "type": "choice",
            "instructions": "Who should handle this?",
            "criteria": ["billing", "technical", "sales"]
        }
    }),
)?;
println!("{}", result.answers["department"]);

Unlike the MLX package this is ported from, inference runs on CPU, CUDA and Metal. Build with --features metal or --features cuda for GPU support.

Re-exports

pub use agent::Agent;

pub use agent::AgentBuilder;

pub use agent::Answer;

pub use agent::Prediction;

pub use config::AgentConfig;

pub use config::EncoderConfig;

pub use model::DecisionModel;

pub use model::ModernBert;

pub use router::ModelName;

pub use router::RouteDecision;

pub use router::Router;

Modules

agent

The public inference runtime. Prompt and result formats follow upstream Laya.

calibrate

Temperature calibration and the confidence metric, ported from upstream.

config

Checkpoint configuration, validated exactly as upstream validates it.

convert

Export a standalone checkpoint with canonical parameter names and a dtype.

email

Email utilities for cleaning and structuring email inputs.

hub

Checkpoint resolution against the Hugging Face Hub.

lang

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

model

ModernBERT encoder and Laya decision heads, in candle.

presets

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

prompt

Prompt construction.

question

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

router

Route a request to the checkpoint best suited to it.

weights

Parameter-name mapping.

Structs

Question

A question after validation, equivalent to upstream’s internal {"t": ..., "ins": ..., "crit": ...} triple.

Tokenizer

A loaded tokenizer plus the four special tokens upstream requires.

Enums

Criteria

Validated criteria, in label order.

Error

Errors raised by the runtime.

QuestionKind

Question type, and the integer the model’s type embedding expects.

Type Aliases

Result