Merge pull request #1981 from AstrBotDevs/feat/r1_filter-integration

feat: 集成r1_filter至框架

astrbot/core/config/default.py

@@ -54,6 +54,7 @@ DEFAULT_CONFIG = {
         "wake_prefix": "",
         "web_search": False,
         "web_search_link": False,
+        "display_reasoning_text": False,
         "identifier": False,
         "datetime_system_prompt": True,
         "default_personality": "default",
@@ -1652,6 +1653,11 @@ CONFIG_METADATA_2 = {
                         "obvious_hint": True,
                         "hint": "开启后，将会传入网页搜索结果的链接给模型，并引导模型输出引用链接。",
                     },
+                    "display_reasoning_text": {
+                        "description": "显示思考内容",
+                        "type": "bool",
+                        "hint": "开启后，将在回复中显示模型的思考过程。",
+                    },
                     "identifier": {
                         "description": "启动识别群员",
                         "type": "bool",

packages/thinking_filter/main.py

@@ -0,0 +1,71 @@
+import re
+from astrbot.api.event import filter, AstrMessageEvent
+from astrbot.api.star import Context, Star, register
+from astrbot.api.provider import LLMResponse
+from openai.types.chat.chat_completion import ChatCompletion
+
+
+@register(
+    "thinking_filter",
+    "Soulter",
+    "可选择是否过滤推理模型的思考内容",
+    "1.0.0",
+    "https://astrbot.app",
+)
+class R1Filter(Star):
+    def __init__(self, context: Context):
+        super().__init__(context)
+        self.display_reasoning_text = (
+            self.context.get_config()
+            .get("provider_settings", {})
+            .get("display_reasoning_text", False)
+        )
+
+    @filter.on_llm_response()
+    async def resp(self, event: AstrMessageEvent, response: LLMResponse):
+        if self.display_reasoning_text:
+            # 显示推理内容的处理逻辑
+            if (
+                response
+                and response.raw_completion
+                and isinstance(response.raw_completion, ChatCompletion)
+                and len(response.raw_completion.choices) > 0
+                and response.raw_completion.choices[0].message
+            ):
+                message = response.raw_completion.choices[0].message
+                reasoning_content = ""  # 初始化 reasoning_content
+
+                # 检查 Groq deepseek-r1-distill-llama-70b 模型的 'reasoning' 属性
+                if hasattr(message, "reasoning") and message.reasoning:
+                    reasoning_content = message.reasoning
+                # 检查 DeepSeek deepseek-reasoner 模型的 'reasoning_content'
+                elif (
+                    hasattr(message, "reasoning_content") and message.reasoning_content
+                ):
+                    reasoning_content = message.reasoning_content
+
+                if reasoning_content:
+                    response.completion_text = (
+                        f"🤔思考：{reasoning_content}\n\n{message.content}"
+                    )
+                else:
+                    response.completion_text = message.content
+        else:
+            # 过滤推理标签的处理逻辑
+            completion_text = response.completion_text
+
+            # 检查并移除 <think> 标签
+            if r"<think>" in completion_text or r"</think>" in completion_text:
+                # 移除配对的标签及其内容
+                completion_text = re.sub(
+                    r"<think>.*?</think>", "", completion_text, flags=re.DOTALL
+                ).strip()
+
+                # 移除可能残留的单个标签
+                completion_text = (
+                    completion_text.replace(r"<think>", "")
+                    .replace(r"</think>", "")
+                    .strip()
+                )
+
+            response.completion_text = completion_text