"""Validate model output without treating imperfect metadata as a job failure."""
from __future__ import annotations

import copy
import json
import re

from knowledge_processing import redact


class ModelOutputError(ValueError):
    """Safe local reason and usage only; never retains the provider response."""

    def __init__(self, reason, input_chars=0, output_chars=0):
        super().__init__(reason)
        self.input_chars = input_chars
        self.output_chars = output_chars


def normalize_output(candidate, text, input_chars):
    output_chars = len(text) if isinstance(text, str) else 0

    def invalid(reason):
        raise ModelOutputError(reason, input_chars, output_chars)

    if not isinstance(text, str) or not text.strip():
        invalid('模型未返回知识正文')
    raw = text.strip().lstrip('\ufeff')
    fence = re.fullmatch(r'```(?:json)?\s*([\s\S]*?)\s*```', raw, flags=re.IGNORECASE)
    if fence:
        raw = fence.group(1).strip()
    try:
        data = json.loads(raw)
    except (ValueError, RecursionError):
        invalid('模型输出不是完整 JSON')
    if not isinstance(data, dict):
        invalid('模型输出不是 JSON 对象')

    # Never stringify objects, invent an answer, or truncate clinical/business facts.
    result = copy.deepcopy(candidate)
    for field, label, maximum in (('question', '问题', 4000), ('answer', '答案', 8000)):
        value = data.get(field)
        if not isinstance(value, str) or not value.strip():
            invalid(f'模型{label}字段缺失、为空或类型不正确')
        value = redact(value)
        if not value or len(value) > maximum:
            invalid(f'模型{label}字段为空或超过 {maximum} 字符')
        result[field] = value

    flags = result.setdefault('flags', [])
    title = data.get('title')
    if not isinstance(title, str) or not title.strip():
        result['title'] = result['question'][:80]
        flags.append('模型未提供有效标题，已从问题生成标题，请核对')
    else:
        title = redact(title)
        result['title'] = title[:160]
        if len(title) > 160:
            flags.append('模型标题过长，已缩短标题；问题与答案未截断')

    category = data.get('category')
    category = redact(category) if isinstance(category, str) else ''
    result['category'] = category if 1 <= len(category) <= 100 else '待分类'
    if result['category'] == '待分类':
        flags.append('模型分类缺失或无效，请人工分类')

    conditions = data.get('conditions')
    conditions = redact(conditions) if isinstance(conditions, str) else ''
    result['conditions'] = conditions if len(conditions) <= 2000 else ''
    if not result['conditions']:
        # Empty conditions deliberately block approval until an editor supplies them.
        flags.append('模型未提供有效适用条件，请核对来源并补充后再审核')

    kind = data.get('kind')
    kind = kind.strip().lower() if isinstance(kind, str) else ''
    aliases = {'问答': 'qa', '标准问答': 'qa', '流程': 'procedure', '业务流程': 'procedure',
               '案例': 'case', '对话案例': 'case'}
    kind = aliases.get(kind, kind)
    result['kind'] = kind if kind in {'qa', 'procedure', 'case'} else 'qa'
    if kind not in {'qa', 'procedure', 'case'}:
        flags.append('模型知识类型无效，暂按问答保留，请核对')
    flags.append('模型整理草稿，须逐项核对来源，不能直接作为正确答案')
    return result, input_chars, output_chars


def original_draft(candidate, error):
    """Keep complete redacted evidence, requiring a human edit before approval."""
    result = copy.deepcopy(candidate)
    result['conditions'] = ''
    result.setdefault('flags', []).append(
        f'模型格式异常：{error}；已保留原始脱敏问答，请人工整理并补充适用条件，不能直接发布')
    return result