#!/usr/bin/env python3
# SPDX-License-Identifier: MIT
# Copyright (c) 2026 何瑜 (OpenGEO)
"""从公开数据复算 OpenGEO 观测记录 #001 的关键数字。

数据：Hugging Face 数据集 Maodian/opengeo-observation-001，固定版本
0ef3c619d1a879c2d848101f459b8512c3c4c548（文件 SHA-256 见下方 FILES）。
文章：https://opengeo.page/observations/001-suzhou-yimei/

只用 Python 3.8+ 标准库。用法：
  python3 recompute.py              # 从 Hugging Face 下载固定版本
  python3 recompute.py --data DIR   # 读本地目录 DIR（里面是 turns.jsonl 等文件）

每一项打印「复算值」和「应为」（文章或数据卡写的值），一致为 OK，不一致为 DIFF。
任何一项 DIFF 或文件哈希不符时，退出码为 1。
"""
from __future__ import annotations

import argparse
import hashlib
import json
import sys
import urllib.request
from collections import Counter
from pathlib import Path

REVISION = "0ef3c619d1a879c2d848101f459b8512c3c4c548"
BASE = f"https://huggingface.co/datasets/Maodian/opengeo-observation-001/resolve/{REVISION}/data/"
FILES = {
    "turns.jsonl": "bd8bed846ab7328f111e84c22eb9369194771e32608b92b35b996e19164a0391",
    "queries.jsonl": "6c4b84d9e78c30a583a1b810ab75af6dca04d15f6174294c48819ec2683ed927",
    "sources.jsonl": "1bd637c89ce5ca559ad1ff5fd960c13b3ecaad666870f69883df828aa3b2fc50",
    "tool_calls.jsonl": "f357af86ec3595731b72c367c1ebec7c2f3cd5f8089d07bea2fed7e742d15cfa",
    "signals.jsonl": "4d8e967de20a4f1f1d983de1e3dd7f897173a1317a9dbb82a43d88e81f203feb",
}


def load(data_dir: Path | None) -> dict[str, list[dict]]:
    out = {}
    for name, digest in FILES.items():
        if data_dir:
            raw = (data_dir / name).read_bytes()
        else:
            with urllib.request.urlopen(BASE + name, timeout=60) as resp:
                raw = resp.read()
        got = hashlib.sha256(raw).hexdigest()
        if got != digest:
            sys.exit(f"{name}: SHA-256 {got} 与固定版本不符，停止。")
        out[name.split(".")[0]] = [json.loads(line) for line in raw.decode("utf-8").splitlines() if line.strip()]
    return out


def pick(rows, platform, session="main", turn=3, **kw):
    return [r for r in rows if r["platform"] == platform and r["session"] == session and r["turn"] == turn
            and all(r.get(k) == v for k, v in kw.items())]


def signal(rows, platform, name, session="main", turn=3):
    vals = [r["value"] for r in pick(rows, platform, session, turn) if r["signal"] == name]
    return vals[0] if vals else None


def main() -> int:
    ap = argparse.ArgumentParser(description=__doc__.splitlines()[0])
    ap.add_argument("--data", type=Path, help="本地数据目录（不联网）")
    d = load(ap.parse_args().data)
    Q, S, C, G = d["queries"], d["sources"], d["tool_calls"], d["signals"]
    checks: list[tuple[str, object, object]] = []

    def check(label, got, want):
        checks.append((label, got, want))

    # 1. 位置
    perms = [r for r in pick(G, "doubao") if r["signal"].startswith("local_permission:")]
    check("豆包第3问 定位权限状态项数", len(perms), 3)
    check("豆包第3问 请求里是否带坐标", signal(G, "doubao", "coordinates_in_request"), "false")
    check("千问 user_city（主会话第1问）", signal(G, "qwen", "user_city", turn=1), "苏州市")
    check("千问第3问 意图 uvl_intent.content", signal(G, "qwen", "uvl_intent.content"), "本地生活-地点周边信息查询")
    yb = sorted(pick(C, "yuanbao"), key=lambda r: r["seq"])
    check("元宝第3问 第1个工具调用", yb[0]["tool_name"] if yb else None, "region_search")

    # 2. 搜索词改写
    dq = pick(Q, "doubao", query_source="search_query_result_block.queries")
    check("豆包第3问 搜索词条数", len(dq), 5)
    check("豆包第3问 第2条搜索词", next((r["query"] for r in dq if r["query_index"] == 2), None), "苏州姑苏XX医美诊所介绍")
    check("Kimi第3问 搜索词条数", len(pick(Q, "kimi")), 3)
    qw = pick(Q, "qwen", query_source="intent_data.hotspot_intent.rewrite_content")
    check("千问第3问 hotspot 改写词条数", len(qw), 1)
    check("千问第3问 hotspot 改写词", qw[0]["query"] if qw else None, "苏州当地医美机构及医疗美容医院")

    # 3. 来源筛选
    check("千问第3问 material_stat.total", signal(G, "qwen", "hotspot_intent.material_stat.total"), "20")
    check("千问第3问 material_stat.sent_to_model", signal(G, "qwen", "hotspot_intent.material_stat.sent_to_model"), "5")

    d3 = pick(S, "doubao", stage="displayed_reference")
    cards = [r for r in d3 if r["source_type"].startswith("card_without_url")]
    linked = [r for r in d3 if r.get("url") or r.get("title")]
    masked = [r for r in d3 if r.get("entity_masked")]
    urls = [r for r in d3 if r.get("url")]
    check("豆包第3问 参考总行数（页面「参考 29 篇」）", len(d3), 29)
    check("豆包第3问 无链接视频卡片", len(cards), 12)
    check("豆包第3问 有标题或链接的来源", len(linked), 17)
    check("豆包第3问 同名诊所相关（entity_masked，网址已删）", len(masked), 6)
    check("豆包第3问 公开数据里可核对的网址", len(urls), 11)
    hosts = Counter(r["domain"] for r in urls)
    check("  其中 抖音 www.iesdouyin.com", hosts["www.iesdouyin.com"], 4)
    check("  其中 医院官网 www.wzrm-hospital.com", hosts["www.wzrm-hospital.com"], 2)
    check("  其中 中华网 m.tech.china.com", hosts["m.tech.china.com"], 2)
    others = sum(v for k, v in hosts.items() if k not in ("www.iesdouyin.com", "www.wzrm-hospital.com", "m.tech.china.com"))
    check("  其中 其他（杭州网、潮新闻、小荷）", others, 3)
    d1 = pick(S, "doubao", turn=1, stage="displayed_reference")
    check("豆包第1问 参考总行数", len(d1), 29)
    check("豆包第1问 short_video 行数", sum(r["source_type"] == "short_video" for r in d1), 10)

    k3 = [r for r in pick(S, "kimi") if r.get("cited")]
    check("Kimi第3问 被引用的不同网页（cited=true 的不同 URL）", len({r["url"] for r in k3}), 8)
    check("Kimi第3问 引用次数（cite_events 合计）", sum(r.get("cite_events", 0) for r in k3), 12)
    check("Kimi第3问 带 :~:text= 段落锚点的引用", sum(r.get("cite_with_text_fragment", 0) for r in k3), 12)
    check("Kimi第3问 signals.citation_events", signal(G, "kimi", "citation_events"), "12")
    check("Kimi第3问 signals.unique_cited_urls", signal(G, "kimi", "unique_cited_urls"), "8")
    check("Kimi第3问 web_search 调用次数", sum(r["tool_name"] == "web_search" for r in pick(C, "kimi")), 1)

    for rnd, (nq, nopen) in {1: (5, 7), 2: (4, 3)}.items():
        check(f"DeepSeek第4问 第{rnd}轮 搜索词", len(pick(Q, "deepseek", "main_q4", 4, round=rnd)), nq)
        check(f"DeepSeek第4问 第{rnd}轮 打开网页（TOOL_OPEN）",
              sum(r["tool_name"] == "TOOL_OPEN" for r in pick(C, "deepseek", "main_q4", 4, round=rnd)), nopen)
    check("DeepSeek第4问 第1轮 搜索返回结果", len(pick(S, "deepseek", "main_q4", 4, round=1, stage="retrieved")), 30)

    bad = 0
    width = max(len(label) for label, _, _ in checks)
    print(f"数据版本 {REVISION[:12]}，文件哈希校验通过\n")
    for label, got, want in checks:
        ok = str(got) == str(want)
        bad += not ok
        print(f"{'OK  ' if ok else 'DIFF'} {label.ljust(width)}  复算 {got}  应为 {want}")
    print(f"\n{len(checks)} 项，{len(checks) - bad} 项一致，{bad} 项不一致")
    print("四家最终推荐的机构数（16/12/10/6 等）不在数据集里，无法复算。")
    return 1 if bad else 0


if __name__ == "__main__":
    sys.exit(main())
