Crate laya 

Rust inference for Laya, a non-autoregressive typed-decision model: a ModernBERT-large encoder plus an RL-trained decision head.

Give it a state (text or JSON) and a set of typed questions; it returns typed answers with calibrated probabilities in a single forward pass. It never generates text.

use laya::{Agent, Question};
use serde_json::json;

let agent = Agent::from_dir("models/laya-base", Default::default())?;
let answers = agent.system_one(
    &json!("My card was charged twice for the same order."),
    &[("department".into(), Question::choice("Which team owns this?", ["billing", "technical", "sales"]))]
        .into_iter()
        .collect(),
)?;
println!("{}", serde_json::to_string_pretty(&answers)?);

Re-exports

pub use config::AgentConfig;

pub use config::EncoderConfig;

pub use question::QType;

pub use question::Question;

Modules

config

Runtime config: the encoder architecture plus the agent’s own decoding settings.

model

The decision model: a frozen-architecture ModernBERT backbone plus the from-scratch decision head (DecisionModel in rl_common.py).

question

Typed questions and their rendering into an encoder sequence.

Structs

Agent

Meta

Options

How to place and run the model.

PromptView

The exact sequence the encoder sees for one question, for inspection and debugging.

Response

TokenView

One token of the rendered sequence.

Usage

Enums

Answer

One typed answer. The variant matches the question’s type.

Segment

Which part of the rendered sequence a token belongs to.