Laya Jev Compatible Server

Yes. The digest shows the official typesafe_sdk client pointed at the local base URL with a test key and zero retries working for all three question types.

The process exits with code 2 when weights are absent, so run the prepare script to fetch models/laya-multilingual before starting the server.

Yes. Bind address, port, checkpoint directory and torch device come from LAYA_HOST, LAYA_PORT, LAYA_ML_DIR and LAYA_DEVICE environment variables.

Run bash scripts/ci.sh from the repo root; protocol details live in docs/server.md alongside the curl and SDK examples.

Python 3.12 or newer, run from the repository root: uv sync, then bash scripts/prepare.sh to fetch the models/laya-multilingual checkpoint, then uv run layalib, which binds 127.0.0.1:8000 by default. The environment is LAYA_HOST, LAYA_PORT, LAYA_ML_DIR for the checkpoint and LAYA_DEVICE for the torch device, so a different port is LAYA_PORT=8001 uv run layalib. It exposes POST /v1/systemone and GET /v1/models.

The SDK example uses api_key="test" because the local server does not authenticate. The protocol is documented in docs/server.md, and the digest shows noul, choice and score questions going through the official typesafe_sdk with RetryPolicy(max_retries=0), reading answers back as r.nouls["billing"].noul, r.choices["tone"].choice and r.scores["urgency"].score. bash scripts/ci.sh is the documented verification path. With LAYA_DEVICE=cpu by default, expect it to be slow.

The three Jev question types against a state you supply. The curl example posts a state object holding a customer document, a model name, and three questions at once: a noul asking whether the ticket is about billing, a choice across calm, frustrated and angry for tone, and a score on a can wait, this week, today rubric for urgency. All three are answered in one request, and the Python SDK example builds the same three through Noul, Choice and Score objects against the local base URL.