"""Classify prompts with the ONNX export (ONNX Runtime; no PyTorch at inference). pip install "laya[onnx]" huggingface_hub python predict_onnx.py "Summarize this: the council voted to expand bike lanes" `laya.onnx_agent.ONNXAgent` reproduces the exact input layout (question header, option markers, state) and temperature scaling the graph was trained with. To call the graph directly, see its `_encode_state` / `_decode_answers`. The Router app's Rust port (crates/router-core/src/router/laya.rs) is a second reference implementation. """ import json import sys from huggingface_hub import snapshot_download from laya.onnx_agent import ONNXAgent REPO = "[org]/router-laya-v1" local = snapshot_download(REPO, allow_patterns=["onnx/*", "tokenizer/*", "rl_agent_config.json", "questions.json", "encoder/*"]) agent = ONNXAgent(local, onnx_path=f"{local}/onnx/router.onnx") questions = json.load(open(f"{local}/questions.json")) for prompt in sys.argv[1:] or ["hi there", "Write a haiku about rain"]: a = agent.predict(prompt, questions)["answers"] print(f"{a['category']['choice']:<18} complexity={a['complexity']['score']:.2f} {prompt[:60]}")