"""Gudao Laya lesson. Default: validate only. --run: download/load real model.

Python 3.10+. Actual model runtime requires the official `laya` package.
No synthetic model answers or fallback predictions are supplied.
"""
import argparse
import importlib.metadata
import json
import math
import os
from pathlib import Path
import sys
import time

MODEL = 'convaiinnovations/laya'
LABELS = ('concept', 'debug', 'clarify')
DEFAULT = {
    'state': '我不懂向量，为什么一句话可以变成一组数字？',
    'questions': {'help_route': {
        'type': 'choice',
        'instructions': '为这条学习求助选择一个最合适的入口。',
        'criteria': {
            'concept': '请求解释概念，没有具体程序错误。',
            'debug': '请求排查代码或具体程序报错。',
            'clarify': '信息不足、与学习无关，或无法明确归入前两类，需要先询问。',
        },
    }},
    'expected': 'concept',
}


def validate_request(request):
    if not isinstance(request, dict):
        raise ValueError('题目必须是 JSON 对象。')
    state = request.get('state')
    if not isinstance(state, str) or not state.strip() or len(state) > 300:
        raise ValueError('state 必须是 1～300 字的求助。')
    if request.get('expected') not in LABELS:
        raise ValueError('expected 必须是 concept/debug/clarify。')
    questions = request.get('questions')
    if not isinstance(questions, dict) or set(questions) != {'help_route'}:
        raise ValueError('本课只接受 help_route 问题。')
    question = questions['help_route']
    if not isinstance(question, dict) or question.get('type') != 'choice':
        raise ValueError('本课问题类型必须是 choice。')
    instruction = question.get('instructions')
    criteria = question.get('criteria')
    if not isinstance(instruction, str) or not instruction.strip() or len(instruction) > 300:
        raise ValueError('请填写不超过 300 字的 instructions。')
    if not isinstance(criteria, dict) or set(criteria) != set(LABELS):
        raise ValueError('criteria 必须含三个课程选项。')
    if any(not isinstance(v, str) or not v.strip() or len(v) > 300 for v in criteria.values()):
        raise ValueError('每个选项需有 1～300 字的说明。')
    return request


def probability(value):
    return type(value) in (int, float) and math.isfinite(value) and 0 <= value <= 1


def validate_result(result):
    if not isinstance(result, dict) or not isinstance(result.get('answers'), dict):
        raise ValueError('SDK 未返回 answers 对象。')
    answer = result['answers'].get('help_route')
    if not isinstance(answer, dict) or answer.get('choice') not in LABELS:
        raise ValueError('模型返回了未知选项。')
    probs = answer.get('probabilities')
    if not isinstance(probs, dict) or set(probs) != set(LABELS):
        raise ValueError('模型返回的概率选项不完整。')
    if not all(probability(v) for v in probs.values()) or abs(sum(probs.values()) - 1) > .002:
        raise ValueError('模型返回的概率分布无效。')
    if probs[answer['choice']] + .0001 < max(probs.values()):
        raise ValueError('模型选项与最大概率不一致。')
    if not probability(answer.get('confidence')):
        raise ValueError('模型返回的 confidence 无效。')
    return answer


def run_model(request, device):
    # This import and all model/network activity occur only after explicit --run.
    os.environ.setdefault('USE_TF', '0')
    import laya
    start = time.perf_counter()
    agent = laya.load(MODEL, subfolder='multilingual', device=device)
    loaded = time.perf_counter()
    result = agent.predict(request['state'], request['questions'])
    finished = time.perf_counter()
    answer = validate_result(result)
    return {
        'status': 'real_model_result',
        'model': MODEL, 'subfolder': 'multilingual',
        'sdk_version': importlib.metadata.version('laya'),
        'device_requested': device, 'device_actual': str(agent.device),
        'load_seconds': round(loaded - start, 4),
        'inference_seconds': round(finished - loaded, 4),
        'request': request, 'result': result,
        'matches_your_expected': answer['choice'] == request['expected'],
        'note': '单题匹配仅核对你的预期，不代表总体准确率。模型未在本次推理中训练。',
    }


def main(argv=None):
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument('--input', type=Path, help='网页下载的 laya-request.json')
    parser.add_argument('--output', type=Path, help='保存报告；默认只打印')
    parser.add_argument('--run', action='store_true', help='下载/加载真实 Laya 并推理')
    parser.add_argument('--device', choices=['cpu', 'cuda'], default='cpu')
    args = parser.parse_args(argv)
    try:
        if args.output and args.output.exists():
            raise ValueError('报告文件已存在，请改用新的 --output 文件名。')
        if args.input and args.input.stat().st_size > 20000:
            raise ValueError('本课题目文件不能超过 20 KB。')
        request = validate_request(json.loads(args.input.read_text(encoding='utf-8-sig')) if args.input else DEFAULT)
        report = run_model(request, args.device) if args.run else {
            'status': 'prepared_only', 'request': request,
            'note': '仅检查题目，没有加载模型、联网或产生预测。加 --run 才会运行真实模型。',
        }
        content = json.dumps(report, ensure_ascii=False, indent=2, allow_nan=False)
        if args.output:
            with args.output.open('x', encoding='utf-8') as stream:
                stream.write(content + '\n')
        print(content)
        return 0
    except Exception as exc:
        print(f'实验停止：{exc}\n没有生成替代或模拟的模型答案。', file=sys.stderr)
        return 1


if __name__ == '__main__':
    if hasattr(sys.stdout, 'reconfigure'):
        sys.stdout.reconfigure(encoding='utf-8')
        sys.stderr.reconfigure(encoding='utf-8')
    raise SystemExit(main())