sycommon-python-lib 0.1.57b5__tar.gz → 0.1.57b7__tar.gz

{sycommon_python_lib-0.1.57b5 → sycommon_python_lib-0.1.57b7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sycommon-python-lib
-Version: 0.1.57b5
+Version: 0.1.57b7
 Summary: Add your description here
 Requires-Python: >=3.11
 Description-Content-Type: text/markdown
@@ -25,6 +25,7 @@ Requires-Dist: pyyaml>=6.0.3
 Requires-Dist: sentry-sdk[fastapi]>=2.49.0
 Requires-Dist: sqlalchemy[asyncio]>=2.0.45
 Requires-Dist: starlette>=0.50.0
+Requires-Dist: tiktoken>=0.12.0
 Requires-Dist: uvicorn>=0.40.0
 
 # sycommon-python-lib

{sycommon_python_lib-0.1.57b5 → sycommon_python_lib-0.1.57b7}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "sycommon-python-lib"
-version = "0.1.57b5"
+version = "0.1.57b7"
 description = "Add your description here"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -26,6 +26,7 @@ dependencies = [
     "sentry-sdk[fastapi]>=2.49.0",
     "sqlalchemy[asyncio]>=2.0.45",
     "starlette>=0.50.0",
+    "tiktoken>=0.12.0",
     "uvicorn>=0.40.0",
 ]
 

{sycommon_python_lib-0.1.57b5 → sycommon_python_lib-0.1.57b7}/src/sycommon/llm/embedding.py

@@ -286,9 +286,12 @@ class Embedding(metaclass=SingletonMeta):
         for i in range(0, len(corpus), batch_size):
             batch_texts = corpus[i: i + batch_size]
 
+            SYLogger.info(
+                f"Requesting embeddings for text: {len(batch_texts)} items (model: {actual_model}, timeout: {timeout or 'None'})")
+
             # 给每个异步任务传入模型名称和超时配置
             tasks = [self._get_embeddings_http_async(
-                text, model=model, timeout=request_timeout) for text in batch_texts]
+                text, model=actual_model, timeout=request_timeout) for text in batch_texts]
             results = await asyncio.gather(*tasks)
 
             for result in results:
@@ -345,9 +348,11 @@ class Embedding(metaclass=SingletonMeta):
         actual_model = model or self.default_reranker_model
         SYLogger.info(
             f"Requesting reranker for top_results: {top_results} (model: {actual_model}, max_concurrency: {self.max_concurrency}, timeout: {timeout or 'None'})")
-
+        # 打印请求参数
+        SYLogger.info(
+            f"Requesting reranker for top_results: {top_results} (model: {actual_model}) (query: {query}) (timeout: {timeout or 'None'})")
         data = await self._get_reranker_http_async(
-            top_results, query, model=model, timeout=request_timeout)
+            top_results, query, model=actual_model, timeout=request_timeout)
         SYLogger.info(
             f"Reranker for top_results completed (model: {actual_model})")
         return data

{sycommon_python_lib-0.1.57b5 → sycommon_python_lib-0.1.57b7}/src/sycommon/llm/get_llm.py

@@ -3,11 +3,15 @@ from langchain.chat_models import init_chat_model
 from sycommon.config.LLMConfig import LLMConfig
 from sycommon.llm.sy_langfuse import LangfuseInitializer
 from sycommon.llm.usage_token import LLMWithAutoTokenUsage
+from typing import Any
 
 
 def get_llm(
     model: str = None,
-    streaming: bool = False
+    *,
+    streaming: bool = False,
+    temperature: float = 0.1,
+    **kwargs: Any
 ) -> LLMWithAutoTokenUsage:
     if not model:
         model = "Qwen2.5-72B"
@@ -16,22 +20,25 @@ def get_llm(
     if not llmConfig:
         raise Exception(f"无效的模型配置：{model}")
 
-    # 初始化Langfuse
+    # 初始化 Langfuse
     langfuse_callbacks, langfuse = LangfuseInitializer.get()
-
     callbacks = [LLMLogger()] + langfuse_callbacks
 
-    llm = init_chat_model(
-        model_provider=llmConfig.provider,
-        model=llmConfig.model,
-        base_url=llmConfig.baseUrl,
-        api_key="-",
-        temperature=0.1,
-        streaming=streaming,
-        callbacks=callbacks
-    )
+    init_params = {
+        "model_provider": llmConfig.provider,
+        "model": llmConfig.model,
+        "base_url": llmConfig.baseUrl,
+        "api_key": "-",
+        "callbacks": callbacks,
+        "temperature": temperature,
+        "streaming": streaming,
+    }
+
+    init_params.update(kwargs)
+
+    llm = init_chat_model(**init_params)
 
     if llm is None:
         raise Exception(f"初始化原始LLM实例失败：{model}")
 
-    return LLMWithAutoTokenUsage(llm, langfuse)
+    return LLMWithAutoTokenUsage(llm, langfuse, llmConfig)

{sycommon_python_lib-0.1.57b5 → sycommon_python_lib-0.1.57b7}/src/sycommon/llm/struct_token.py

@@ -1,21 +1,109 @@
+import tiktoken
 from typing import Dict, List, Optional, Any
 from langfuse import Langfuse, LangfuseSpan, propagate_attributes
 from sycommon.llm.llm_logger import LLMLogger
 from langchain_core.runnables import Runnable, RunnableConfig
-from langchain_core.messages import BaseMessage, HumanMessage
+from langchain_core.messages import BaseMessage, SystemMessage, HumanMessage
 from sycommon.llm.llm_tokens import TokensCallbackHandler
 from sycommon.logging.kafka_log import SYLogger
+from sycommon.config.LLMConfig import LLMConfig
 from sycommon.tools.env import get_env_var
 from sycommon.tools.merge_headers import get_header_value
 
 
 class StructuredRunnableWithToken(Runnable):
-    """带Token统计的Runnable类"""
+    """
+    统一功能 Runnable：Trace追踪 + Token统计 + 自动上下文压缩
+    """
 
-    def __init__(self, retry_chain: Runnable, langfuse: Optional[Langfuse]):
+    def __init__(
+        self,
+        retry_chain: Runnable,
+        langfuse: Optional[Langfuse] = None,
+        llmConfig: Optional[LLMConfig] = None,
+        model_name: str = "Qwen2.5-72B",
+        enable_compression: bool = True,
+        threshold_ratio: float = 0.8
+    ):
         super().__init__()
         self.retry_chain = retry_chain
         self.langfuse = langfuse
+        self.llmConfig = llmConfig
+        self.model_name = model_name
+        self.enable_compression = enable_compression
+        self.threshold_ratio = threshold_ratio
+
+        # 初始化 Tokenizer
+        try:
+            self.encoding = tiktoken.encoding_for_model(model_name)
+        except KeyError:
+            self.encoding = tiktoken.get_encoding("cl100k_base")
+
+    def _count_tokens(self, messages: List[BaseMessage]) -> int:
+        """快速估算 Token 数量"""
+        num_tokens = 0
+        for message in messages:
+            num_tokens += 4  # 每条消息的固定开销
+            # 兼容 content 是字符串或者 dict 的情况
+            content = message.content
+            if isinstance(content, str):
+                num_tokens += len(self.encoding.encode(content))
+            elif isinstance(content, list):  # 多模态或复杂结构
+                for item in content:
+                    if isinstance(item, dict) and "text" in item:
+                        num_tokens += len(self.encoding.encode(item["text"]))
+            elif isinstance(content, dict):
+                num_tokens += len(self.encoding.encode(str(content)))
+        return num_tokens
+
+    async def _acompress_context(self, messages: List[BaseMessage]) -> List[BaseMessage]:
+        """执行异步上下文压缩"""
+        # 策略：保留 System Prompt + 最近 N 条，中间的摘要
+        keep_last_n = 4
+
+        # 分离系统消息和对话消息
+        system_msgs = [m for m in messages if isinstance(m, SystemMessage)]
+        conversation = [
+            m for m in messages if not isinstance(m, SystemMessage)]
+
+        if len(conversation) <= keep_last_n:
+            return messages
+
+        to_summarize = conversation[:-keep_last_n]
+        keep_recent = conversation[-keep_last_n:]
+
+        # 构造摘要 Prompt
+        # 注意：这里直接使用 retry_chain 进行摘要，防止死循环
+        summary_prompt = [
+            SystemMessage(
+                content="请将上下文内容进行摘要，保留关键信息，将内容压缩到原来长度的50%左右，保留关键信息。"),
+            HumanMessage(content=f"历史记录:\n{to_summarize}\n\n摘要:")
+        ]
+
+        try:
+            SYLogger.info(
+                f"🚀 Triggering compression: {len(to_summarize)} messages -> summary")
+
+            # 调用子链生成摘要
+            # 【关键】必须清空 callbacks，否则 Langfuse 会递归追踪，导致死循环或噪音
+            summary_result = await self.retry_chain.ainvoke(
+                {"messages": summary_prompt},
+                config=RunnableConfig(callbacks=[])
+            )
+
+            summary_text = summary_result.content if hasattr(
+                summary_result, 'content') else str(summary_result)
+
+            # 重组消息：System + Summary + Recent
+            new_messages = system_msgs + \
+                [SystemMessage(
+                    content=f"[History Summary]: {summary_text}")] + keep_recent
+            return new_messages
+
+        except Exception as e:
+            SYLogger.error(
+                f"❌ Compression failed: {e}, using original context.")
+            return messages
 
     def _adapt_input(self, input: Any) -> List[BaseMessage]:
         """适配输入格式"""
@@ -25,6 +113,10 @@ class StructuredRunnableWithToken(Runnable):
             return [input]
         elif isinstance(input, str):
             return [HumanMessage(content=input)]
+        elif isinstance(input, dict) and "messages" in input:
+            # 如果已经是标准格式字典，直接提取
+            msgs = input["messages"]
+            return msgs if isinstance(msgs, list) else [msgs]
         elif isinstance(input, dict) and "input" in input:
             return [HumanMessage(content=str(input["input"]))]
         else:
@@ -40,7 +132,7 @@ class StructuredRunnableWithToken(Runnable):
         token_handler = TokensCallbackHandler()
 
         if config is None:
-            processed_config = {"callbacks": [], "metadata": {}}
+            processed_config = RunnableConfig(callbacks=[], metadata={})
         else:
             processed_config = config.copy()
             if "callbacks" not in processed_config:
@@ -59,6 +151,7 @@ class StructuredRunnableWithToken(Runnable):
             callbacks.append(LLMLogger())
         callbacks.append(token_handler)
 
+        # 去重
         callback_types = {}
         unique_callbacks = []
         for cb in callbacks:
@@ -131,6 +224,8 @@ class StructuredRunnableWithToken(Runnable):
                 user_id=user_id
             )
 
+            # 【同步模式下不建议触发压缩，因为压缩本身是异步调用 LLM】
+            # 如果同步也要压缩，需要用 asyncio.run(...)，这里暂时保持原逻辑直接透传
             adapted_input = self._adapt_input(input)
             input_data = {"messages": adapted_input}
 
@@ -169,12 +264,26 @@ class StructuredRunnableWithToken(Runnable):
                 user_id=user_id
             )
 
+            # 1. 适配输入
             adapted_input = self._adapt_input(input)
+
+            # 2. 检查并执行上下文压缩 (仅在异步模式且开启时)
+            if self.enable_compression:
+                max_tokens = self.llmConfig.maxTokens
+                current_tokens = self._count_tokens(adapted_input)
+
+                if current_tokens > max_tokens * self.threshold_ratio:
+                    SYLogger.warning(
+                        f"⚠️ Context limit reached: {current_tokens}/{max_tokens}")
+                    # 执行压缩，替换 adapted_input
+                    adapted_input = await self._acompress_context(adapted_input)
+
             input_data = {"messages": adapted_input}
 
             if span:
                 span.update_trace(input=input_data)
 
+            # 3. 调用子链
             structured_result = await self.retry_chain.ainvoke(
                 input_data,
                 config=processed_config

{sycommon_python_lib-0.1.57b5 → sycommon_python_lib-0.1.57b7}/src/sycommon/llm/usage_token.py

@@ -6,6 +6,7 @@ from langchain_core.output_parsers import PydanticOutputParser
 from langchain_core.messages import BaseMessage, HumanMessage
 from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
 from pydantic import BaseModel, ValidationError, Field
+from sycommon.config.LLMConfig import LLMConfig
 from sycommon.llm.struct_token import StructuredRunnableWithToken
 
 
@@ -13,9 +14,10 @@ class LLMWithAutoTokenUsage(BaseChatModel):
     """自动为结构化调用返回token_usage的LLM包装类"""
     llm: BaseChatModel = Field(default=None)
     langfuse: Optional[Langfuse] = Field(default=None, exclude=True)
+    llmConfig: Optional[LLMConfig] = Field(default=None, exclude=True)
 
-    def __init__(self, llm: BaseChatModel, langfuse: Langfuse, **kwargs):
-        super().__init__(llm=llm, langfuse=langfuse, **kwargs)
+    def __init__(self, llm: BaseChatModel, langfuse: Langfuse, llmConfig: LLMConfig, **kwargs):
+        super().__init__(llm=llm, langfuse=langfuse, llmConfig=llmConfig, **kwargs)
 
     def with_structured_output(
         self,
@@ -106,7 +108,7 @@ class LLMWithAutoTokenUsage(BaseChatModel):
                 "initial": 0.1, "max": 3.0, "exp_base": 2.0, "jitter": 1.0}
         )
 
-        return StructuredRunnableWithToken(retry_chain, self.langfuse)
+        return StructuredRunnableWithToken(retry_chain, self.langfuse, self.llmConfig)
 
     # ========== 实现BaseChatModel抽象方法 ==========
     def _generate(self, messages, stop=None, run_manager=None, ** kwargs):

{sycommon_python_lib-0.1.57b5 → sycommon_python_lib-0.1.57b7}/src/sycommon/rabbitmq/rabbitmq_client.py

@@ -117,112 +117,112 @@ class RabbitMQClient:
             logger.info(f"队列重建成功: {self.queue_name}")
 
     async def connect(self) -> None:
+        """连接方法（修复恢复消费失效问题）"""
         if self._closed:
             raise RuntimeError("客户端已关闭，无法重新连接")
 
-        # 1. 并发控制：使用 _connect_lock 保证只有一个协程在执行连接流程
-        async with self._connect_lock:
-            # 如果已经在连了，等待其完成
+        # 1. 获取 Condition 锁
+        await self._connect_condition.acquire()
+
+        try:
+            # ===== 阶段 A: 快速检查与等待 =====
+            if await self.is_connected:
+                self._connect_condition.release()
+                return
+
             if self._connecting:
-                logger.debug("连接正在进行中，等待现有连接完成...")
                 try:
-                    # 等待条件变量，超时设为 60 秒防止死等
-                    await asyncio.wait_for(
-                        self._connect_condition.wait_for(
-                            lambda: not self._connecting),
-                        timeout=60.0
-                    )
+                    logger.debug("连接正在进行中，等待现有连接完成...")
+                    await asyncio.wait_for(self._connect_condition.wait(), timeout=60.0)
                 except asyncio.TimeoutError:
+                    self._connect_condition.release()
                     raise RuntimeError("等待连接超时")
 
-                # 等待结束后，再次检查状态
-                if not await self.is_connected:
+                if await self.is_connected:
+                    self._connect_condition.release()
+                    return
+                else:
+                    self._connect_condition.release()
                     raise RuntimeError("等待重连后，连接状态依然无效")
-                return
 
-            # 标记开始连接
+            # ===== 阶段 B: 标记开始连接 =====
             self._connecting = True
+            # 【关键】释放锁，允许其他协程进入等待逻辑
+            self._connect_condition.release()
 
-        # 释放 _connect_lock，允许其他协程读取状态，但在连接完成前阻止新的连接请求
-        # 注意：这里释放了 _connect_lock，但 self._connecting = True 阻止了新的连接流程
+        except Exception as e:
+            if self._connect_condition.locked():
+                self._connect_condition.release()
+            raise
 
+        # === 阶段 C: 执行耗时的连接逻辑 (此时已释放锁，不阻塞其他协程) ===
         try:
-            # --- 阶段1: 清理旧资源 ---
-            # 重新获取锁进行资源清理
-            async with self._connect_lock:
-                was_consuming = self._consumer_tag is not None
-
-                if self._channel_conn and self._conn_close_callback:
-                    try:
-                        self._channel_conn.close_callbacks.discard(
-                            self._conn_close_callback)
-                    except Exception:
-                        pass
-
-                self._channel = None
-                self._channel_conn = None
-                self._exchange = None
-                self._queue = None
-                self._conn_close_callback = None
-
-            # --- 阶段2: 获取新连接 (耗时IO) ---
+            # --- 步骤 1: 记录旧状态并清理资源 ---
+            # 必须在清理前记录状态
+            was_consuming = self._consumer_tag is not None
+
+            # 清理连接回调，防止旧的连接关闭触发新的重连
+            if self._channel_conn:
+                try:
+                    if self._channel_conn.close_callbacks:
+                        self._channel_conn.close_callbacks.clear()
+                except Exception:
+                    pass
+
+            # 统一重置资源状态
+            self._channel = None
+            self._channel_conn = None
+            self._exchange = None
+            self._queue = None
+            self._consumer_tag = None
+
+            # --- 步骤 2: 获取新连接 ---
             self._channel, self._channel_conn = await self.connection_pool.acquire_channel()
 
-            # 设置回调
+            # 设置连接关闭回调
             def on_conn_closed(conn, exc):
-                logger.warning(f"检测到连接关闭: {exc}")
+                logger.warning(f"检测到底层连接关闭: {exc}")
                 if not self._closed and not self._connecting:
                     asyncio.create_task(self._safe_reconnect())
 
-            self._conn_close_callback = on_conn_closed
             if self._channel_conn:
-                self._channel_conn.close_callbacks.add(
-                    self._conn_close_callback)
+                self._channel_conn.close_callbacks.add(on_conn_closed)
 
-            # 重建资源
+            # --- 步骤 3: 重建基础资源 (交换机和队列) ---
             await self._rebuild_resources()
 
-            # --- 阶段3: 恢复消费 ---
-            if was_consuming and self._message_handler and self.queue_name and self.queue_name.endswith(f".{self.app_name}"):
-                logger.info("🔄 检测到重连前处于消费状态，尝试自动恢复...")
+            # --- 步骤 4: 恢复消费 ---
+            if was_consuming and self._message_handler:
+                logger.info("🔄 检测到重连前处于消费状态，尝试自动恢复消费...")
                 try:
-                    self._queue = await self._channel.declare_queue(
-                        name=self.queue_name,
-                        durable=self.durable,
-                        auto_delete=self.auto_delete,
-                        passive=False,
-                    )
-                    await self._queue.bind(exchange=self._exchange, routing_key=self.routing_key)
-                    self._consumer_tag = await self._queue.consume(self._process_message_callback)
+                    # 直接调用 start_consuming 来恢复，它内部包含了完整的队列检查和绑定逻辑
+                    self._consumer_tag = await self.start_consuming()
                     logger.info(f"✅ 消费已自动恢复: {self._consumer_tag}")
                 except Exception as e:
                     logger.error(f"❌ 自动恢复消费失败: {e}")
                     self._consumer_tag = None
-            else:
-                self._consumer_tag = None
 
             logger.info("客户端连接初始化完成")
 
         except Exception as e:
             logger.error(f"客户端连接失败: {str(e)}", exc_info=True)
-
-            # 异常时清理资源
-            async with self._connect_lock:
-                if self._channel_conn and self._conn_close_callback:
-                    self._channel_conn.close_callbacks.discard(
-                        self._conn_close_callback)
-                self._channel = None
-                self._channel_conn = None
-                self._consumer_tag = None
-
+            # 异常时彻底清理
+            if self._channel_conn and self._channel_conn.close_callbacks:
+                self._channel_conn.close_callbacks.clear()
+            self._channel = None
+            self._channel_conn = None
+            self._queue = None
+            self._consumer_tag = None
             raise
 
         finally:
-            # 【关键修复】必须在持有 Condition 内部锁的情况下调用 notify_all
-            # 这里使用 async with self._connect_condition: 自动完成 acquire() ... notify_all() ... release()
-            async with self._connect_condition:
+            # === 阶段 D: 恢复状态并通知 ===
+            await self._connect_condition.acquire()
+            try:
                 self._connecting = False
                 self._connect_condition.notify_all()
+            finally:
+                self._connect_condition.release()
 
     async def _safe_reconnect(self):
         """安全重连任务（仅用于被动监听连接关闭）"""
@@ -256,11 +256,6 @@ class RabbitMQClient:
             self._message_handler = handler
 
     async def _process_message_callback(self, message: AbstractIncomingMessage):
-        # 记录消息的原始追踪ID
-        original_trace_id = message.headers.get(
-            "trace-id") if message.headers else None
-        current_retry = 0
-
         try:
             msg_obj: MQMsgModel
 
@@ -278,44 +273,21 @@ class RabbitMQClient:
                     body=message.body.decode("utf-8"),
                     routing_key=message.routing_key,
                     delivery_tag=message.delivery_tag,
-                    traceId=original_trace_id,
+                    traceId=message.headers.get(
+                        "trace-id") if message.headers else SYLogger.get_trace_id(),
                 )
 
-            # 2. 设置日志上下文
-            # 注意：如果 header 中有 x-last-retry-ts，说明之前重试过
-            current_retry = int(message.headers.get("x-retry-count", 0))
             SYLogger.set_trace_id(msg_obj.traceId)
 
             # 3. 执行业务逻辑
             if self._message_handler:
                 await self._message_handler(msg_obj, message)
 
-            # 4. 业务成功，Ack (移除 finally 中的 ack，成功即确认)
             await message.ack()
 
         except Exception as e:
-            logger.error(f"消息处理异常 (第 {current_retry} 次尝试): {e}", exc_info=True)
-
-            # 【核心修复】使用原生 Nack + Requeue
-            if current_retry >= 3:
-                # 超过重试次数，丢弃消息（或进入死信队列）
-                logger.warning(f"重试次数超限 (3次)，丢弃消息: {message.delivery_tag}")
-                await message.reject(requeue=False)
-            else:
-                # 还没到重试上限，重新入队
-                # 为了防止立即重试导致的死循环，我们需要人为增加一点延迟
-                # 但 Nack 本身不支持延迟，所以这里只能快速 Nack 让它尽快回来，
-                # 并在业务层（或外层）做好限流保护。
-
-                # 如果你有延迟队列插件，可以 publish 到延迟交换机。
-                # 如果没有，直接 requeue 是最安全的不丢包方案。
-                logger.info(f"消息处理失败，重新入队等待重试... (当前重试: {current_retry})")
-
-                # 技巧：如果你不想立即重试，可以 Nack(False) 然后手动 Publish 延迟消息
-                # 但为了解决你当前的“死循环”问题，直接 Nack(True) 是最有效的
-                # 延迟5秒
-                await asyncio.sleep(5)
-                await message.nack(requeue=True)
+            logger.error(f"消息处理异常: {e}", exc_info=True)
+            await message.ack()
 
     async def start_consuming(self) -> Optional[ConsumerTag]:
         if self._closed:

{sycommon_python_lib-0.1.57b5 → sycommon_python_lib-0.1.57b7}/src/sycommon_python_lib.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sycommon-python-lib
-Version: 0.1.57b5
+Version: 0.1.57b7
 Summary: Add your description here
 Requires-Python: >=3.11
 Description-Content-Type: text/markdown
@@ -25,6 +25,7 @@ Requires-Dist: pyyaml>=6.0.3
 Requires-Dist: sentry-sdk[fastapi]>=2.49.0
 Requires-Dist: sqlalchemy[asyncio]>=2.0.45
 Requires-Dist: starlette>=0.50.0
+Requires-Dist: tiktoken>=0.12.0
 Requires-Dist: uvicorn>=0.40.0
 
 # sycommon-python-lib

{sycommon_python_lib-0.1.57b5 → sycommon_python_lib-0.1.57b7}/src/sycommon_python_lib.egg-info/requires.txt

@@ -19,4 +19,5 @@ pyyaml>=6.0.3
 sentry-sdk[fastapi]>=2.49.0
 sqlalchemy[asyncio]>=2.0.45
 starlette>=0.50.0
+tiktoken>=0.12.0
 uvicorn>=0.40.0