# -*- coding: utf-8 -*-
"""DeepSeek 审千问: 本地跨模型审核 v1 (2026-08-24 用户令 "建立对千问的审核机制").
分层: ① citation_gate 确定性引文回抓闸 — 治本地审核模型的"引用失实"病 (14b/32b 各 1 例实逮);
     ② r1 审核员只出质疑清单, 不出裁决 (推理无界、结论受闸);
     ③ 审核员的每条质疑自身也过 ① — 摘录在被审文本中找不到 → 该条自动废弃并标记.
能力边界 (gold 实测 2/8): 本地审 = 快筛层; 审级定谳只认云端双模审."""
import json, re, time, urllib.request

URL = 'http://localhost:11434'
REVIEWER_PREF = ['deepseek-r1:14b', 'deepseek-r1:32b']
RULES = ('①答案中的台号/关键数字必须在【事实】里有出处 ②等级词(优秀/良好/报警/危险/不可判)须与事实一致, 无擅自升降级 '
         '③结论方向须与证据方向一致 ④【事实】未提的部件/时间/机制不得出现断言')

PROMPT = """你是独立审核员, 审【答案】是否忠实于【事实】。规则: {rules}
只输出 JSON, 不要任何其他文字: {{"issues": [{{"quote": "答案原文摘录(逐字)", "problem": "一句话问题", "rule": "①-④"}}], "verdict": "PASS 或 N条质疑"}}
没有问题就输出 {{"issues": [], "verdict": "PASS"}}。

【问题】{q}
【事实(工具产出, 权威)】{facts}
【答案(被审)】{answer}
"""


def _turbs(text):
    out = set()
    for a, b, c in re.findall(r'WTG-?(\d{1,2})|(\d{1,2})#|(\d{1,2})号机?', str(text)):
        out.add(int(a or b or c))
    return out


def _nums(text):
    return set(re.findall(r'\d+\.?\d*', str(text)))


def _strip_labels(text):
    """剔非断言型数字: 窗名 wMMDD / ISO 日期 / 版本号 — 它们是标签不是数值断言.
    (2026-08-25 实逮: 30b 答案里 w0127/w0707/w0724/w0805 被判'数字无出处' 8 处全是窗名假阳)"""
    s = str(text)
    s = re.sub(r'[wW]\d{4}\b', ' ', s)                    # 窗名 w0707
    s = re.sub(r'\b\d{4}-\d{2}(-\d{2})?\b', ' ', s)      # ISO 日期
    s = re.sub(r'\bv\d+(\.\d+)*\b', ' ', s)              # 版本号 v5 / v2.0
    return s


def citation_gate(claim, source, num_min_digits=2):
    """claim 里引用的台号/数字必须在 source 有出处; 返回失实清单 (确定性, 零模型)."""
    bad = []
    claim_s, source_s = _strip_labels(claim), _strip_labels(source)
    st, sn = _turbs(source), _nums(source_s)
    norm = lambda x: x.rstrip('0').rstrip('.') if '.' in x else x
    sn_n = sn | {norm(x) for x in sn} | {x.split('.')[0] for x in sn}
    for t in _turbs(claim):
        if t not in st:
            bad.append(f'台号 {t}# 无出处')
    for n in _nums(claim_s):
        if len(n.replace('.', '')) < num_min_digits or n in sn_n:
            continue
        if norm(n) in sn_n or n.split('.')[0] in sn_n:
            continue
        bad.append(f'数字 {n} 无出处')
    return bad


def _chat(prompt, model, npred=800, timeout=180):
    body = {'model': model, 'messages': [{'role': 'user', 'content': prompt}], 'stream': False,
            'think': True, 'keep_alive': '30m', 'options': {'num_predict': npred, 'num_ctx': 16384}}
    req = urllib.request.Request(URL + '/api/chat', data=json.dumps(body).encode('utf-8'), headers={'Content-Type': 'application/json'})
    d = json.loads(urllib.request.urlopen(req, timeout=timeout).read())
    return (d.get('message', {}) or {}).get('content', '')


def pick_reviewer():
    try:
        req = urllib.request.Request(URL + '/api/tags')
        tags = json.loads(urllib.request.urlopen(req, timeout=4).read())
        names = {m['name'] for m in tags.get('models', [])}
        for p in REVIEWER_PREF:
            if p in names:
                return p
    except Exception:
        pass
    return None


def review_answer(question, answer, facts, model=None):
    """跨模型审核: 机器闸 + r1 质疑清单 (质疑自过闸). 返回 dict, 不抛异常."""
    t0 = time.time()
    out = dict(machine_flags=citation_gate(answer, facts), issues=[], discarded=[], verdict='', model=None, seconds=0)
    model = model or pick_reviewer()
    if not model:
        out['verdict'] = ('机器闸: ' + '; '.join(out['machine_flags'])) if out['machine_flags'] else 'PASS(仅机器闸, 审核模型不在线)'
        return out
    out['model'] = model
    try:
        txt = _chat(PROMPT.format(rules=RULES, q=str(question)[:500], facts=str(facts)[:6000], answer=str(answer)[:4000]), model)
        m = re.search(r'\{.*\}', txt, flags=re.S)
        r = json.loads(m.group(0)) if m else {}
        norm = lambda s: re.sub(r'\s+', '', str(s))
        na = norm(answer)
        for it in (r.get('issues') or []):
            q_ = norm(it.get('quote', ''))
            # ③ 审核员摘录必须真在答案里 + 质疑正文自身过引文闸
            if q_ and q_[:40] in na and not citation_gate(it.get('problem', ''), str(answer) + str(facts)):
                out['issues'].append(it)
            else:
                out['discarded'].append(dict(it, reason='引用失实(摘录不在答案中或质疑引了不存在的台号/数字)'))
        n = len(out['issues']) + len(out['machine_flags'])
        out['verdict'] = 'PASS' if n == 0 else f'{n} 条质疑'
    except Exception as e:
        out['verdict'] = f'审核失败: {e}'
    out['seconds'] = round(time.time() - t0, 1)
    return out