better 更好的llm使用统计

src/plugins/chat/config.py

@@ -56,7 +56,6 @@ class BotConfig:
     llm_reasoning: Dict[str, str] = field(default_factory=lambda: {})
     llm_reasoning_minor: Dict[str, str] = field(default_factory=lambda: {})
     llm_normal: Dict[str, str] = field(default_factory=lambda: {})
-    llm_normal_minor: Dict[str, str] = field(default_factory=lambda: {})
     llm_topic_judge: Dict[str, str] = field(default_factory=lambda: {})
     llm_summary_by_topic: Dict[str, str] = field(default_factory=lambda: {})
     llm_emotion_judge: Dict[str, str] = field(default_factory=lambda: {})
@@ -235,7 +234,6 @@ class BotConfig:
                 "llm_reasoning",
                 "llm_reasoning_minor",
                 "llm_normal",
-                "llm_normal_minor",
                 "llm_topic_judge",
                 "llm_summary_by_topic",
                 "llm_emotion_judge",

src/plugins/chat/emoji_manager.py

@@ -38,9 +38,9 @@ class EmojiManager:
 
     def __init__(self):
         self._scan_task = None
-        self.vlm = LLM_request(model=global_config.vlm, temperature=0.3, max_tokens=1000, request_type="image")
+        self.vlm = LLM_request(model=global_config.vlm, temperature=0.3, max_tokens=1000, request_type="emoji")
         self.llm_emotion_judge = LLM_request(
-            model=global_config.llm_emotion_judge, max_tokens=600, temperature=0.8, request_type="image"
+            model=global_config.llm_emotion_judge, max_tokens=600, temperature=0.8, request_type="emoji"
         )  # 更高的温度，更少的token（后续可以根据情绪来调整温度）
 
     def _ensure_emoji_dir(self):
@@ -111,7 +111,7 @@ class EmojiManager:
             if not text_for_search:
                 logger.error("无法获取文本的情绪")
                 return None
-            text_embedding = await get_embedding(text_for_search)
+            text_embedding = await get_embedding(text_for_search, request_type="emoji")
             if not text_embedding:
                 logger.error("无法获取文本的embedding")
                 return None
@@ -310,7 +310,7 @@ class EmojiManager:
                     logger.info(f"[检查] 表情包检查通过: {check}")
 
                 if description is not None:
-                    embedding = await get_embedding(description)
+                    embedding = await get_embedding(description, request_type="emoji")
                     # 准备数据库记录
                     emoji_record = {
                         "filename": filename,

src/plugins/chat/llm_generator.py

@@ -32,10 +32,17 @@ class ResponseGenerator:
             temperature=0.7,
             max_tokens=1000,
             stream=True,
+            request_type="response",
+        )
+        self.model_v3 = LLM_request(
+            model=global_config.llm_normal, temperature=0.7, max_tokens=3000, request_type="response"
+        )
+        self.model_r1_distill = LLM_request(
+            model=global_config.llm_reasoning_minor, temperature=0.7, max_tokens=3000, request_type="response"
+        )
+        self.model_sum = LLM_request(
+            model=global_config.llm_summary_by_topic, temperature=0.7, max_tokens=3000, request_type="relation"
         )
-        self.model_v3 = LLM_request(model=global_config.llm_normal, temperature=0.7, max_tokens=3000)
-        self.model_r1_distill = LLM_request(model=global_config.llm_reasoning_minor, temperature=0.7, max_tokens=3000)
-        self.model_v25 = LLM_request(model=global_config.llm_normal_minor, temperature=0.7, max_tokens=3000)
         self.current_model_type = "r1"  # 默认使用 R1
         self.current_model_name = "unknown model"
 
@@ -175,7 +182,7 @@ class ResponseGenerator:
             """
 
             # 调用模型生成结果
-            result, _, _ = await self.model_v25.generate_response(prompt)
+            result, _, _ = await self.model_sum.generate_response(prompt)
             result = result.strip()
 
             # 解析模型输出的结果

src/plugins/chat/utils.py

@@ -55,9 +55,9 @@ def is_mentioned_bot_in_message(message: MessageRecv) -> bool:
     return False
 
 
-async def get_embedding(text):
+async def get_embedding(text, request_type="embedding"):
     """获取文本的embedding向量"""
-    llm = LLM_request(model=global_config.embedding, request_type="embedding")
+    llm = LLM_request(model=global_config.embedding, request_type=request_type)
     # return llm.get_embedding_sync(text)
     return await llm.get_embedding(text)
 

src/plugins/memory_system/memory.py

@@ -174,9 +174,9 @@ class Memory_graph:
 class Hippocampus:
     def __init__(self, memory_graph: Memory_graph):
         self.memory_graph = memory_graph
-        self.llm_topic_judge = LLM_request(model=global_config.llm_topic_judge, temperature=0.5, request_type="topic")
+        self.llm_topic_judge = LLM_request(model=global_config.llm_topic_judge, temperature=0.5, request_type="memory")
         self.llm_summary_by_topic = LLM_request(
-            model=global_config.llm_summary_by_topic, temperature=0.5, request_type="topic"
+            model=global_config.llm_summary_by_topic, temperature=0.5, request_type="memory"
         )
 
     def get_all_node_names(self) -> list:
@@ -375,6 +375,8 @@ class Hippocampus:
         return topic_num
 
     async def operation_build_memory(self):
+        logger.debug("------------------------------------开始构建记忆--------------------------------------")
+        start_time = time.time()
         memory_samples = self.get_memory_sample()
         all_added_nodes = []
         all_added_edges = []
@@ -426,6 +428,12 @@ class Hippocampus:
         logger.success(f"强化连接: {', '.join(all_added_edges)}")
         # logger.success(f"强化连接: {', '.join(all_added_edges)}")
         self.sync_memory_to_db()
+        
+        end_time = time.time()
+        logger.success(
+            f"--------------------------记忆构建完成：耗时: {end_time - start_time:.2f} "
+            "秒--------------------------"
+        )
 
     def sync_memory_to_db(self):
         """检查并同步内存中的图结构与数据库"""

src/plugins/models/utils_model.py

@@ -581,7 +581,8 @@ class LLM_request:
                         completion_tokens=completion_tokens,
                         total_tokens=total_tokens,
                         user_id="system",  # 可以根据需要修改 user_id
-                        request_type="embedding",  # 请求类型为 embedding
+                        # request_type="embedding",  # 请求类型为 embedding
+                        request_type=self.request_type,  # 请求类型为 text
                         endpoint="/embeddings",  # API 端点
                     )
                     return result["data"][0].get("embedding", None)

template/bot_config_template.toml

@@ -128,52 +128,60 @@ enable = true
 
 
 #下面的模型若使用硅基流动则不需要更改，使用ds官方则改成.env.prod自定义的宏，使用自定义模型则选择定位相似的模型自己填写
-#推理模型：
+#推理模型
+
 [model.llm_reasoning] #回复模型1 主要回复模型
 name = "Pro/deepseek-ai/DeepSeek-R1"
+# name = "Qwen/QwQ-32B"
 provider = "SILICONFLOW"
-pri_in = 0 #模型的输入价格（非必填，可以记录消耗）
+pri_in = 4 #模型的输入价格（非必填，可以记录消耗）
-pri_out = 0 #模型的输出价格（非必填，可以记录消耗）
+pri_out = 16 #模型的输出价格（非必填，可以记录消耗）
 
 [model.llm_reasoning_minor] #回复模型3 次要回复模型
 name = "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B"
 provider = "SILICONFLOW"
+pri_in = 1.26 #模型的输入价格（非必填，可以记录消耗）
+pri_out = 1.26 #模型的输出价格（非必填，可以记录消耗）
 
 #非推理模型
 
 [model.llm_normal] #V3 回复模型2 次要回复模型
 name = "Pro/deepseek-ai/DeepSeek-V3"
 provider = "SILICONFLOW"
+pri_in = 2 #模型的输入价格（非必填，可以记录消耗）
+pri_out = 8 #模型的输出价格（非必填，可以记录消耗）
 
-[model.llm_normal_minor] #V2.5
+[model.llm_emotion_judge] #表情包判断 
-name = "deepseek-ai/DeepSeek-V2.5"
-provider = "SILICONFLOW"
-
-[model.llm_emotion_judge] #主题判断 0.7/m
 name = "Qwen/Qwen2.5-14B-Instruct"
 provider = "SILICONFLOW"
+pri_in = 0.7
+pri_out = 0.7
 
-[model.llm_topic_judge] #主题判断：建议使用qwen2.5 7b
+[model.llm_topic_judge] #记忆主题判断：建议使用qwen2.5 7b
 name = "Pro/Qwen/Qwen2.5-7B-Instruct"
 provider = "SILICONFLOW"
+pri_in = 0
+pri_out = 0
 
-[model.llm_summary_by_topic] #建议使用qwen2.5 32b 及以上
+[model.llm_summary_by_topic] #概括模型，建议使用qwen2.5 32b 及以上
 name = "Qwen/Qwen2.5-32B-Instruct"
 provider = "SILICONFLOW"
-pri_in = 0
+pri_in = 1.26
-pri_out = 0
+pri_out = 1.26
 
-[model.moderation] #内容审核 未启用
+[model.moderation] #内容审核，开发中
 name = ""
 provider = "SILICONFLOW"
-pri_in = 0
+pri_in = 1.0
-pri_out = 0
+pri_out = 2.0
 
 # 识图模型
 
-[model.vlm] #图像识别 0.35/m
+[model.vlm] #图像识别 
-name = "Pro/Qwen/Qwen2-VL-7B-Instruct"
+name = "Pro/Qwen/Qwen2.5-VL-7B-Instruct"
 provider = "SILICONFLOW"
+pri_in = 0.35
+pri_out = 0.35
 
 #嵌入模型