OCR过滤：加多/cmd指令列表碎片检测 + 内容相关性检查防偶然匹配

2026-07-29 15:29:42 +08:00
parent 47bd990f76
commit 659f667d74
2 changed files with 136 additions and 97 deletions
@@ -2,6 +2,7 @@

 import logging
 import threading
+import jieba

 logger = logging.getLogger("msgbot.autoreply")

@@ -23,7 +24,7 @@ class AutoReply:
        logger.info(f"自动回复 {'已开启' if enabled else '已关闭'}")

    def set_min_score(self, score: float):
-        self._min_score = max(0.0, min(1.0, score))
+        self._min_score = max(0.0, min(10.0, score))

    @property
    def enabled(self) -> bool:
@@ -44,6 +45,13 @@ class AutoReply:
                f"[{platform}] 匹配度 {score:.2f} < 阈值 {self._min_score}，跳过回复"
            )
            return None
+        # 内容相关性检查：查询词与匹配文档标题至少有一个共同2字+词，防偶然匹配
+        if not self._has_content_relevance(msg, top_doc):
+            heading = top_doc.get("heading", "")[:40]
+            logger.debug(
+                f"[{platform}] 内容相关性检查失败，跳过: msg={msg[:40]!r} doc_heading={heading!r}"
+            )
+            return None
        with self._lock:
            return self._get_answer(platform, msg, top_doc, score)

@@ -64,6 +72,20 @@ class AutoReply:

        return {"platform": platform, "answer": answer, "sender": sender}

+    @staticmethod
+    def _has_content_relevance(query: str, doc: dict) -> bool:
+        """查询词与匹配文档标题至少有一个共同2字+词，防偶然匹配。
+        短查询(≤2个2字+词)直接放行，不做严格检查。
+        """
+        q_tokens = set(t for t in jieba.cut(query) if len(t) >= 2)
+        if len(q_tokens) <= 2:
+            return True  # 短查询不拦截
+        doc_text = doc.get("heading", "") + " " + doc.get("title", "")
+        d_tokens = set(t for t in jieba.cut(doc_text) if len(t) >= 2)
+        if not (q_tokens & d_tokens):
+            return False
+        return True
+
    def _get_answer(self, platform: str, msg: str, doc: dict, score: float) -> str or None:
        """提取答案 + 防重复检查。
        优先返回 doc['answer']（Q&A 段），兜底使用 doc['content']。
@@ -325,79 +325,49 @@ class MessageMonitor:
    # ── 双击 ESC 急停 ──────────────────────────────────────

    def _esc_check_loop(self):
-        """独立 daemon 线程：底层键盘钩子检测双击 ESC → 强制急停。
+        """独立 daemon 线程：RegisterHotKey 全局热键检测双击 ESC → 强制急停。

-        使用 WH_KEYBOARD_LL 系统级全局钩子，不依赖窗口焦点，
-        任何进程按 ESC 都能捕获（包括游戏全屏、浏览器、桌面等）。
-        钩子回调必须运行在安装钩子的线程消息泵里（GetMessage）。
+        使用 RegisterHotKey 注册全局热键 + GetMessage 消息泵，
+        不依赖窗口焦点，任何进程按 ESC 都能捕获。
+        比 WH_KEYBOARD_LL 更可靠（避免 ctypes 64位调用约定问题）。
        """
        from ctypes import wintypes

-        # ── Win32 常量 & 结构体 ──
-        _WH_KEYBOARD_LL = 13
-        _WM_KEYDOWN = 0x0100
        _VK_ESCAPE = 0x1B
+        _WM_HOTKEY = 0x0312
+        _HOTKEY_ID_ESC = 0x4553  # 'ES'

-        class _KBDLLHOOKSTRUCT(ctypes.Structure):
-            _fields_ = [
-                ("vkCode", wintypes.DWORD),
-                ("scanCode", wintypes.DWORD),
-                ("flags", wintypes.DWORD),
-                ("time", wintypes.DWORD),
-                ("dwExtraInfo", ctypes.POINTER(ctypes.c_ulong)),
-            ]
-
-        _HOOKPROC = ctypes.CFUNCTYPE(
-            ctypes.c_long, ctypes.c_int, wintypes.WPARAM, wintypes.LPARAM
-        )
-
-        def _ll_hook(nCode, wParam, lParam):
-            """底层键盘钩子回调 — 在 _esc_check_loop 线程执行。"""
-            if nCode >= 0 and wParam == _WM_KEYDOWN:
-                kb = ctypes.cast(lParam, ctypes.POINTER(_KBDLLHOOKSTRUCT)).contents
-                if kb.vkCode == _VK_ESCAPE:
-                    now = time.time()
-                    if _ll_hook.last_press > 0 and (now - _ll_hook.last_press) < 0.5:
-                        _ll_hook.last_press = 0.0
-                        logger.warning("⚠️ 双击ESC急停触发！强制终止所有操作...")
-                        self._emergency_stop()
-                    else:
-                        _ll_hook.last_press = now
-            return ctypes.windll.user32.CallNextHookEx(None, nCode, wParam, lParam)
-
-        _ll_hook.last_press = 0.0
-
-        # 引用保持（防止 GC 回收回调 → 系统崩溃）
-        self._esc_hook_proc = _HOOKPROC(_ll_hook)
-
-        # ⭐ 步骤1: 先创建线程消息队列（PeekMessage 强制创建）
-        #   WH_KEYBOARD_LL 要求安装线程有消息队列，否则钩子回调无法投递
+        # ⭐ 步骤1: 创建消息队列
        _msg = wintypes.MSG()
-        ctypes.windll.user32.PeekMessageW(ctypes.byref(_msg), None, 0, 0, 0)  # PM_NOREMOVE
+        ctypes.windll.user32.PeekMessageW(ctypes.byref(_msg), None, 0, 0, 0)

-        # ⭐ 步骤2: 安装钩子 — hmod 必须为 NULL (0)
-        #   GetModuleHandleW(None) 返回 Python.exe 的模块句柄 ≠ NULL → 钩子静默失败！
-        hook_handle = ctypes.windll.user32.SetWindowsHookExW(
-            _WH_KEYBOARD_LL, self._esc_hook_proc, None, 0  # hmod = NULL
-        )
-        if not hook_handle:
+        # ⭐ 步骤2: 注册全局热键（hwnd=NULL → 投递到线程消息队列）
+        #   MOD_NOREPEAT = 0x4000：长按不重复触发，只有按下时触发一次
+        ok = ctypes.windll.user32.RegisterHotKey(None, _HOTKEY_ID_ESC, 0x4000, _VK_ESCAPE)
+        if not ok:
            err = ctypes.windll.kernel32.GetLastError()
-            logger.error(
-                "[急停] 底层键盘钩子安装失败 (err=%d) — ESC 急停不可用", err
-            )
+            # err=1409 = ERROR_HOTKEY_ALREADY_REGISTERED，其他热键已占 ESC
+            logger.error("[急停] RegisterHotKey 失败 (err=%d) — ESC 急停不可用", err)
            return
-        logger.info("[急停] 底层键盘钩子安装成功，ESC 急停已激活")
+        logger.info("[急停] 全局热键注册成功，双击 ESC 急停已激活")

-        # ⭐ 步骤3: 消息泵 — WH_KEYBOARD_LL 通过 GetMessage 分发钩子回调
+        # ⭐ 步骤3: 消息泵 — 检测双击 ESC
+        last_esc_time = 0.0
        while self._running:
            ret = ctypes.windll.user32.GetMessageW(ctypes.byref(_msg), None, 0, 0)
            if ret <= 0:
-                break  # WM_QUIT 或出错
-            ctypes.windll.user32.TranslateMessage(ctypes.byref(_msg))
-            ctypes.windll.user32.DispatchMessageW(ctypes.byref(_msg))
+                break
+            if _msg.message == _WM_HOTKEY and _msg.wParam == _HOTKEY_ID_ESC:
+                now = time.time()
+                if last_esc_time > 0 and (now - last_esc_time) < 0.5:
+                    last_esc_time = 0.0
+                    logger.warning("⚠️ 双击ESC急停触发！强制终止所有操作...")
+                    self._emergency_stop()
+                else:
+                    last_esc_time = now

-        # 线程退出 → 卸载钩子
-        ctypes.windll.user32.UnhookWindowsHookEx(hook_handle)
+        # 线程退出 → 注销热键
+        ctypes.windll.user32.UnregisterHotKey(None, _HOTKEY_ID_ESC)

    def _emergency_stop(self):
        """强制停止监控+杀死 EasyOCR+清空发送锁。"""
@@ -1417,55 +1387,63 @@ class MessageMonitor:
            logger.debug(f"[ocr] stage[{i}] split -> {len(parts)} parts: "
                         f"{[p[:60] for p in parts if p.strip()]}")
            buf = ''
+            last_ts = None  # 跟踪上一时间戳，用于判断消息块年龄
            for part in parts:
                if re.match(r'^\d{1,2}:\d{1,2}', part):
                    if buf.strip():
-                        if self._is_markdown_document(buf.strip()):
-                            extracted = self._extract_user_text_from_md_fragment(buf.strip())
-                            for e in extracted:
-                                if e:
-                                    # 再检：提取后的文本不应是纯 md 语法
-                                    if self._is_markdown_document(e):
-                                        logger.debug(f"[ocr] stage[{i}] md-only after extraction: {e[:60]!r}")
-                                        continue
-                                    expanded.append(e)
-                            if not extracted:
-                                logger.debug(f"[ocr] stage[{i}] markdown fragment skipped: {buf.strip()[:80]!r}")
+                        if last_ts and self._is_timestamp_old(last_ts, 300):
+                            logger.debug(f"[ocr] stage[{i}] 时间戳{last_ts}过旧({300}s)，跳过: {buf.strip()[:60]!r}")
                        else:
-                            msg = re.sub(r'\s*(?:[A-Za-z0-9]~)+\s*$', '', buf.strip())
-                            if msg:
-                                msg = re.sub(r'\s+#[\u4e00-\u9fff][^\s#\n]{0,30}$', '', msg).strip()
+                            if self._is_markdown_document(buf.strip()):
+                                extracted = self._extract_user_text_from_md_fragment(buf.strip())
+                                for e in extracted:
+                                    if e:
+                                        # 再检：提取后的文本不应是纯 md 语法
+                                        if self._is_markdown_document(e):
+                                            logger.debug(f"[ocr] stage[{i}] md-only after extraction: {e[:60]!r}")
+                                            continue
+                                        expanded.append(e)
+                                if not extracted:
+                                    logger.debug(f"[ocr] stage[{i}] markdown fragment skipped: {buf.strip()[:80]!r}")
+                            else:
+                                msg = re.sub(r'\s*(?:[A-Za-z0-9]~)+\s*$', '', buf.strip())
                                if msg:
-                                    if self._is_reply_fragment(msg):
-                                        logger.debug(f"[ocr] stage[{i}] 回复残片跳过: {msg[:60]!r}")
-                                    else:
-                                        expanded.append(msg)
+                                    msg = re.sub(r'\s+#[\u4e00-\u9fff][^\s#\n]{0,30}$', '', msg).strip()
+                                    if msg:
+                                        if self._is_reply_fragment(msg):
+                                            logger.debug(f"[ocr] stage[{i}] 回复残片跳过: {msg[:60]!r}")
+                                        else:
+                                            expanded.append(msg)
                    buf = ''
+                    last_ts = part
                else:
                    buf += part
            if buf.strip():
-                if self._is_markdown_document(buf.strip()):
-                    extracted = self._extract_user_text_from_md_fragment(buf.strip())
-                    for e in extracted:
-                        if e:
-                            if self._is_markdown_document(e):
-                                logger.debug(f"[ocr] stage[{i}] md-only after extraction: {e[:60]!r}")
-                                continue
-                            if self._is_reply_fragment(e):
-                                logger.debug(f"[ocr] stage[{i}] md提取回复残片跳过: {e[:60]!r}")
-                                continue
-                            expanded.append(e)
-                    if not extracted:
-                        logger.debug(f"[ocr] stage[{i}] markdown fragment skipped: {buf.strip()[:80]!r}")
+                if last_ts and self._is_timestamp_old(last_ts, 300):
+                    logger.debug(f"[ocr] stage[{i}] 时间戳{last_ts}过旧({300}s)，跳过尾部: {buf.strip()[:60]!r}")
                else:
-                    msg = re.sub(r'\s*(?:[A-Za-z0-9]~)+\s*$', '', buf.strip())
-                    if msg:
-                        msg = re.sub(r'\s+#[\u4e00-\u9fff][^\s#\n]{0,30}$', '', msg).strip()
+                    if self._is_markdown_document(buf.strip()):
+                        extracted = self._extract_user_text_from_md_fragment(buf.strip())
+                        for e in extracted:
+                            if e:
+                                if self._is_markdown_document(e):
+                                    logger.debug(f"[ocr] stage[{i}] md-only after extraction: {e[:60]!r}")
+                                    continue
+                                if self._is_reply_fragment(e):
+                                    logger.debug(f"[ocr] stage[{i}] md提取回复残片跳过: {e[:60]!r}")
+                                    continue
+                                expanded.append(e)
+                        if not extracted:
+                            logger.debug(f"[ocr] stage[{i}] markdown fragment skipped: {buf.strip()[:80]!r}")
+                    else:
+                        msg = re.sub(r'\s*(?:[A-Za-z0-9]~)+\s*$', '', buf.strip())
                        if msg:
-                            if self._is_reply_fragment(msg):
-                                logger.debug(f"[ocr] stage[{i}] 回复残片跳过: {msg[:60]!r}")
-                            else:
-                                expanded.append(msg)
+                            msg = re.sub(r'\s+#[\u4e00-\u9fff][^\s#\n]{0,30}$', '', msg).strip()
+                            if msg:
+                                if self._is_reply_fragment(msg):
+                                    logger.debug(f"[ocr] stage[{i}] 回复残片跳过: {msg[:60]!r}")
+                                else:
+                                    expanded.append(msg)

        # —— 逐行过滤 ——
        ocr_lines = expanded
@@ -1620,6 +1598,14 @@ class MessageMonitor:
                logger.debug(f"[{platform}] OCR 过滤: 管道符碎片(绿色气泡) line={line!r}")
                continue

+            # —— 游戏指令列表碎片：重复 /cmd 模式 ——
+            # OCR 截到命令帮助文本如 "/sdfl主指令 /sdfl reload /sdfl update"
+            # 正常用户消息极少含 ≥2 个独立 /command
+            slash_cmds = re.findall(r'(?:^|\s)/(\w{2,})', line)
+            if len(slash_cmds) >= 2:
+                logger.debug(f"[{platform}] OCR 过滤: 指令列表碎片(多/cmd={slash_cmds}) line={line!r}")
+                continue
+
            # —— JSON/YAML 碎片兜底：OCR 截屏混入配置面板 ——
            # 行首是 "key": value 或 key: value（YAML）→ 配置面板碎片
            if re.match(r'^["\']?[a-zA-Z_][\w]*["\']?\s*:\s*["\d\[\{tfn]', line):
@@ -2170,6 +2156,37 @@ class MessageMonitor:
                    return True
        return False

+    @staticmethod
+    def _is_timestamp_old(ts_str: str, max_age: int = 300) -> bool:
+        """检查 HH:MM 或 HH:MM:SS 时间戳是否超过 max_age 秒。
+        处理跨午夜边界：如果 diff > 12h 说明是昨天的时间戳。
+        返回 True = 时间戳太老，应跳过。
+        """
+        ts = ts_str.strip()
+        if not ts:
+            return False
+        m = re.match(r'^(\d{1,2}):(\d{2})(?::(\d{2}))?$', ts)
+        if not m:
+            return False
+        h, mi, s = int(m.group(1)), int(m.group(2)), int(m.group(3) or 0)
+        if h > 23 or mi > 59 or s > 59:
+            return False
+        now = datetime.now()
+        try:
+            parsed = now.replace(hour=h, minute=mi, second=s, microsecond=0)
+        except ValueError:
+            return False
+        diff = (now - parsed).total_seconds()
+        # 跨午夜: diff > 12h 说明时间戳来自昨天
+        if diff > 43200:
+            parsed = parsed.replace(day=parsed.day - 1)
+            diff = (now - parsed).total_seconds()
+        # 未来时间戳: diff < -12h 说明是明天（当前 23:55，时间戳 00:05）
+        if diff < -43200:
+            parsed = parsed.replace(day=parsed.day + 1)
+            diff = (now - parsed).total_seconds()
+        return diff > max_age
+
    @staticmethod
    def _is_system_message(text: str) -> bool:
        """过滤机器人/系统消息，不作为用户消息处理。"""