feat(ai): clean reasoning content from replies

2026-04-07 09:23:48 +08:00
parent 496463c442
commit 51fe971cda
4 changed files with 73 additions and 10 deletions
--- a/plugins/dify/main.py
+++ b/plugins/dify/main.py
@@ -21,7 +21,7 @@ from utils.revoke.message_auto_revoke import MessageAutoRevoke
 from utils.robot_cmd.robot_command import Feature, PermissionStatus, GroupBotManager
 from utils.decorator.points_decorator import plugin_points_cost
 from utils.media_downloader import MediaDownloader
-from utils.string_utils import remove_trailing_content, remove_grok_render_tags
+from utils.string_utils import remove_reasoning_content, remove_trailing_content, remove_grok_render_tags
 from wechat_ipad import WechatAPIClient
 from wechat_ipad.models.message import MessageType
 import aiohttp
@@ -250,6 +250,12 @@ class DifyPlugin(MessagePluginInterface):
                             response: str, roomid: str) -> Tuple[bool, str]:
        """发送响应消息的辅助方法"""
        try:
+            if response and not os.path.isfile(response):
+                response = remove_reasoning_content(response)
+                response = remove_trailing_content(response)
+                response = remove_grok_render_tags(response)
+                response = re.sub(r'\n{3,}', '\n\n', response).strip()
+
            # 判断是否为本地文件路径
            if os.path.isfile(response):
                # 如果是文件路径，使用发送文件方法
@@ -549,6 +555,12 @@ class DifyPlugin(MessagePluginInterface):
                # 获取token使用情况
                total_tokens = response_data.get("data", {}).get("total_tokens", 0)

+                if answer and not os.path.isfile(answer):
+                    answer = remove_reasoning_content(answer)
+                    answer = remove_trailing_content(answer)
+                    answer = remove_grok_render_tags(answer)
+                    answer = re.sub(r'\n{3,}', '\n\n', answer).strip()
+
                # 更新会话历史
                self.conversations[session_id].append({
                    "role": "user",
--- a/plugins/message_summary/config.toml
+++ b/plugins/message_summary/config.toml
@@ -6,7 +6,10 @@ enabled = true
 [api]
 api_key = "app-McGLzBhBjeBCSEi7n83MtuTo"
 api_url = "http://192.168.2.240/v1/chat-messages"
-response_mode = "streaming"
+response_mode = "blocking"
+connect_timeout_seconds = 10
+request_timeout_seconds = 180
+retry_delays_seconds = [10, 20]

 [output]
 output_dir = "output"
--- a/plugins/message_summary/main.py
+++ b/plugins/message_summary/main.py
@@ -21,7 +21,7 @@ from utils.decorator.rate_limit_decorator import group_feature_rate_limit
 from utils.markdown_to_image import convert_md_str_to_image
 from utils.revoke.message_auto_revoke import MessageAutoRevoke
 from utils.robot_cmd.robot_command import GroupBotManager, PermissionStatus
-from utils.string_utils import remove_trailing_content
+from utils.string_utils import remove_reasoning_content, remove_trailing_content
 from utils.wechat.contact_manager import ContactManager
 from utils.wechat.message_to_db import MessageStorage
 from wechat_ipad import WechatAPIClient
@@ -84,7 +84,10 @@ class MessageSummaryPlugin(MessagePluginInterface):
            api_config = self._config.get("api", {})
            self._api_key = api_config.get("api_key", "app-McGLzBhBjeBCSEi7n83MtuTo")
            self._api_url = api_config.get("api_url", "http://192.168.2.240/v1/chat-messages")
-            self._response_mode = api_config.get("response_mode", "streaming")
+            self._response_mode = api_config.get("response_mode", "blocking")
+            self._connect_timeout_seconds = int(api_config.get("connect_timeout_seconds", 10))
+            self._request_timeout_seconds = int(api_config.get("request_timeout_seconds", 180))
+            self._retry_delays_seconds = api_config.get("retry_delays_seconds", [10, 20])
            self.message_storage = MessageStorage()
            db_manager = context.get("db_manager")
            if db_manager:
@@ -275,6 +278,16 @@ class MessageSummaryPlugin(MessagePluginInterface):
        tokens_info = f"\n\n【tokens】输入: {prompt_tokens} 生成: {completion_tokens} 总: {total_tokens}"
        return answer + tokens_info

+    def _clean_summary_output(self, answer: str) -> str:
+        """清理总结输出中的思考内容和无关尾部内容"""
+        if not answer:
+            return answer
+
+        cleaned = remove_reasoning_content(answer)
+        cleaned = remove_trailing_content(cleaned)
+        cleaned = re.sub(r'\n{3,}', '\n\n', cleaned).strip()
+        return cleaned
+
    def _get_revoke_manager(self) -> Optional[MessageAutoRevoke]:
        """优先使用消息上下文中的撤回器，定时任务场景则懒初始化一个"""
        if self.revoke:
@@ -357,12 +370,15 @@ class MessageSummaryPlugin(MessagePluginInterface):
            "Accept": "text/event-stream" if self._response_mode == "streaming" else "application/json"
        }

-        max_retries = 3
-        retry_delays = [2, 4]
+        max_retries = len(self._retry_delays_seconds) + 1

        for attempt in range(1, max_retries + 1):
            try:
-                custom_timeout = ClientTimeout(total=None, connect=10, sock_read=300)
+                custom_timeout = ClientTimeout(
+                    total=None,
+                    connect=self._connect_timeout_seconds,
+                    sock_read=self._request_timeout_seconds
+                )
                conn = aiohttp.TCPConnector(keepalive_timeout=60)  # 保持连接活跃
                async with aiohttp.ClientSession(connector=conn, timeout=custom_timeout) as session:
                    async with session.post(self._api_url, headers=headers, json=data) as response:
@@ -381,8 +397,7 @@ class MessageSummaryPlugin(MessagePluginInterface):

                        # 提取回答内容
                        answer = response_data.get("answer", "")
-                        # 去除广告内容pollinations.ai 的广告
-                        # answer = remove_trailing_content(answer)
+                        answer = self._clean_summary_output(answer)
                        spath = ""
                        # 提取token使用情况
                        metadata = response_data.get("metadata", {})
@@ -420,7 +435,7 @@ class MessageSummaryPlugin(MessagePluginInterface):
                self.LOG.error(f"处理总结时出现未知错误: attempt={attempt}/{max_retries}, error={e}")

            if attempt < max_retries:
-                delay = retry_delays[attempt - 1] if attempt - 1 < len(retry_delays) else retry_delays[-1]
+                delay = self._retry_delays_seconds[attempt - 1]
                self.LOG.warning(f"群总结生成失败，准备重试: attempt={attempt}/{max_retries}, delay={delay}s")
                await asyncio.sleep(delay)

--- a/utils/string_utils.py
+++ b/utils/string_utils.py
@@ -22,3 +22,36 @@ def remove_grok_render_tags(text: str) -> str:
    cleaned = re.sub(r'<\s*grok:[^>]+?>[\s\S]*?<\s*/\s*grok:[^>]+?>', '', text, flags=re.IGNORECASE)
    cleaned = re.sub(r'<\s*grok:[^>]+?/?>', '', cleaned, flags=re.IGNORECASE)
    return cleaned
+
+
+def remove_reasoning_content(text: str) -> str:
+    """移除模型返回中的思考/推理内容，仅保留最终可展示结果。"""
+    if not text:
+        return text
+
+    cleaned = text
+
+    # 常见的思考标签
+    cleaned = re.sub(r'<\s*think\s*>[\s\S]*?<\s*/\s*think\s*>', '', cleaned, flags=re.IGNORECASE)
+    cleaned = re.sub(r'<\s*thinking\s*>[\s\S]*?<\s*/\s*thinking\s*>', '', cleaned, flags=re.IGNORECASE)
+    cleaned = re.sub(r'<\s*reasoning\s*>[\s\S]*?<\s*/\s*reasoning\s*>', '', cleaned, flags=re.IGNORECASE)
+
+    # 常见 markdown 思考段落
+    cleaned = re.sub(
+        r'^\s*(#{1,6}\s*)?(思考过程|推理过程|分析过程|reasoning|thinking)\s*[:：]?\s*$[\s\S]*?(?=^\s*(#{1,6}\s*)?\S|\Z)',
+        '',
+        cleaned,
+        flags=re.IGNORECASE | re.MULTILINE
+    )
+
+    # 某些模型会输出“思考内容：...最终答案：...”
+    cleaned = re.sub(
+        r'^\s*(思考内容|思维链|推理过程|分析过程)\s*[:：][\s\S]*?(?=(最终答案|总结|摘要|#|\Z))',
+        '',
+        cleaned,
+        flags=re.IGNORECASE
+    )
+
+    cleaned = remove_grok_render_tags(cleaned)
+    cleaned = re.sub(r'\n{3,}', '\n\n', cleaned).strip()
+    return cleaned