sycommon-python-lib 0.1.57b2__tar.gz → 0.1.57b3__tar.gz

{sycommon_python_lib-0.1.57b2 → sycommon_python_lib-0.1.57b3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sycommon-python-lib
-Version: 0.1.57b2
+Version: 0.1.57b3
 Summary: Add your description here
 Requires-Python: >=3.11
 Description-Content-Type: text/markdown

{sycommon_python_lib-0.1.57b2 → sycommon_python_lib-0.1.57b3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "sycommon-python-lib"
-version = "0.1.57b2"
+version = "0.1.57b3"
 description = "Add your description here"
 readme = "README.md"
 requires-python = ">=3.11"

sycommon_python_lib-0.1.57b3/src/sycommon/llm/embedding.py

@@ -0,0 +1,368 @@
+import asyncio
+import aiohttp
+import atexit
+from typing import Union, List, Optional, Dict
+from sycommon.config.Config import SingletonMeta
+from sycommon.config.EmbeddingConfig import EmbeddingConfig
+from sycommon.config.RerankerConfig import RerankerConfig
+from sycommon.logging.kafka_log import SYLogger
+
+
+class Embedding(metaclass=SingletonMeta):
+    def __init__(self):
+        # 1. 并发限制
+        self.max_concurrency = 20
+        # 保留默认模型名称
+        self.default_embedding_model = "bge-large-zh-v1.5"
+        self.default_reranker_model = "bge-reranker-large"
+
+        # 初始化默认模型的基础URL
+        self.embeddings_base_url = EmbeddingConfig.from_config(
+            self.default_embedding_model).baseUrl
+        self.reranker_base_url = RerankerConfig.from_config(
+            self.default_reranker_model).baseUrl
+
+        # [修复] 缓存配置URL，避免高并发下重复读取配置文件
+        self._embedding_url_cache: Dict[str, str] = {
+            self.default_embedding_model: self.embeddings_base_url
+        }
+        self._reranker_url_cache: Dict[str, str] = {
+            self.default_reranker_model: self.reranker_base_url
+        }
+
+        # [修复] 缓存模型的向量维度，用于生成兜底零向量
+        self._model_dim_cache: Dict[str, int] = {}
+
+        # 并发信号量
+        self.semaphore = asyncio.Semaphore(self.max_concurrency)
+        self.default_timeout = aiohttp.ClientTimeout(total=None)
+
+        # 核心优化：创建全局可复用的ClientSession（连接池复用）
+        self.session = None
+        # 重试配置（可根据需要调整）
+        self.max_retry_attempts = 3  # 最大重试次数
+        self.retry_wait_base = 0.5   # 基础等待时间（秒）
+
+        # [修复] 注册退出钩子，确保程序结束时关闭连接池
+        atexit.register(self._sync_close_session)
+
+    async def init_session(self):
+        """初始化全局ClientSession（仅创建一次）"""
+        if self.session is None or self.session.closed:
+            # 配置连接池参数，适配高并发
+            connector = aiohttp.TCPConnector(
+                limit=self.max_concurrency * 2,  # 连接池最大连接数（建议是并发数的2倍）
+                limit_per_host=self.max_concurrency,  # 每个域名的最大连接数
+                ttl_dns_cache=300,  # DNS缓存时间
+                enable_cleanup_closed=True  # 自动清理关闭的连接
+            )
+            self.session = aiohttp.ClientSession(
+                connector=connector,
+                timeout=self.default_timeout
+            )
+
+    async def close_session(self):
+        """关闭全局Session（程序退出时调用）"""
+        if self.session and not self.session.closed:
+            await self.session.close()
+
+    def _sync_close_session(self):
+        """同步关闭Session的封装，供atexit调用"""
+        # 注意：atexit在主线程运行，如果当前没有事件循环，这个操作可能会受限
+        # 但它能捕获大多数正常退出的场景。对于asyncio程序，建议显式调用cleanup
+        try:
+            loop = asyncio.get_event_loop()
+            if loop.is_running():
+                # 如果loop还在跑，创建一个任务去关闭
+                loop.create_task(self.close_session())
+            else:
+                # 如果loop已经停止，尝试运行一次
+                loop.run_until_complete(self.close_session())
+        except Exception:
+            # 静默处理清理失败，避免退出报错
+            pass
+
+    async def _retry_request(self, func, *args, **kwargs):
+        """
+        原生异步重试封装函数
+        Args:
+            func: 待重试的异步函数
+            *args: 函数参数
+            **kwargs: 函数关键字参数
+        Returns:
+            函数执行结果，重试失败返回None
+        """
+        attempt = 0
+        while attempt < self.max_retry_attempts:
+            try:
+                return await func(*args, **kwargs)
+            except (aiohttp.ClientConnectionResetError, asyncio.TimeoutError, aiohttp.ClientError) as e:
+                attempt += 1
+                if attempt >= self.max_retry_attempts:
+                    SYLogger.error(
+                        f"Request failed after {attempt} retries: {str(e)}")
+                    return None
+                # 指数退避等待：0.5s → 1s → 2s（最大不超过5s）
+                wait_time = min(self.retry_wait_base * (2 ** (attempt - 1)), 5)
+                SYLogger.warning(
+                    f"Retry {func.__name__} (attempt {attempt}/{self.max_retry_attempts}): {str(e)}, wait {wait_time}s")
+                await asyncio.sleep(wait_time)
+            except Exception as e:
+                # 非重试类异常直接返回None
+                SYLogger.error(
+                    f"Non-retryable error in {func.__name__}: {str(e)}")
+                return None
+        return None
+
+    def _get_embedding_url(self, model: str) -> str:
+        """获取Embedding URL（带缓存）"""
+        if model not in self._embedding_url_cache:
+            self._embedding_url_cache[model] = EmbeddingConfig.from_config(
+                model).baseUrl
+        return self._embedding_url_cache[model]
+
+    def _get_reranker_url(self, model: str) -> str:
+        """获取Reranker URL（带缓存）"""
+        if model not in self._reranker_url_cache:
+            self._reranker_url_cache[model] = RerankerConfig.from_config(
+                model).baseUrl
+        return self._reranker_url_cache[model]
+
+    async def _get_embeddings_http_core(
+        self,
+        input: Union[str, List[str]],
+        encoding_format: str = None,
+        model: str = None,
+        timeout: aiohttp.ClientTimeout = None,
+        **kwargs
+    ):
+        """embedding请求核心逻辑（剥离重试，供重试封装调用）"""
+        await self.init_session()  # 确保Session已初始化
+        async with self.semaphore:
+            request_timeout = timeout or self.default_timeout
+            target_model = model or self.default_embedding_model
+
+            # [修复] 使用缓存获取URL
+            target_base_url = self._get_embedding_url(target_model)
+            url = f"{target_base_url}/v1/embeddings"
+
+            request_body = {
+                "model": target_model,
+                "input": input,
+                "encoding_format": encoding_format or "float"
+            }
+            request_body.update(kwargs)
+
+            # 复用全局Session
+            async with self.session.post(
+                url,
+                json=request_body,
+                timeout=request_timeout
+            ) as response:
+                if response.status != 200:
+                    error_detail = await response.text()
+                    SYLogger.error(
+                        f"Embedding request failed (model: {target_model}): {error_detail}")
+                    return None
+                return await response.json()
+
+    async def _get_embeddings_http_async(
+        self,
+        input: Union[str, List[str]],
+        encoding_format: str = None,
+        model: str = None,
+        timeout: aiohttp.ClientTimeout = None, ** kwargs
+    ):
+        """对外暴露的embedding请求方法（包含重试）"""
+        return await self._retry_request(
+            self._get_embeddings_http_core,
+            input, encoding_format, model, timeout, ** kwargs
+        )
+
+    async def _get_reranker_http_core(
+        self,
+        documents: List[str],
+        query: str,
+        top_n: Optional[int] = None,
+        model: str = None,
+        max_chunks_per_doc: Optional[int] = None,
+        return_documents: Optional[bool] = True,
+        return_len: Optional[bool] = True,
+        timeout: aiohttp.ClientTimeout = None, ** kwargs
+    ):
+        """reranker请求核心逻辑（剥离重试，供重试封装调用）"""
+        await self.init_session()  # 确保Session已初始化
+        async with self.semaphore:
+            request_timeout = timeout or self.default_timeout
+            target_model = model or self.default_reranker_model
+
+            # [修复] 使用缓存获取URL
+            target_base_url = self._get_reranker_url(target_model)
+            url = f"{target_base_url}/v1/rerank"
+
+            request_body = {
+                "model": target_model,
+                "documents": documents,
+                "query": query,
+                "top_n": top_n or len(documents),
+                "max_chunks_per_doc": max_chunks_per_doc,
+                "return_documents": return_documents,
+                "return_len": return_len,
+            }
+            request_body.update(kwargs)
+
+            # 复用全局Session
+            async with self.session.post(
+                url,
+                json=request_body,
+                timeout=request_timeout
+            ) as response:
+                if response.status != 200:
+                    error_detail = await response.text()
+                    SYLogger.error(
+                        f"Rerank request failed (model: {target_model}): {error_detail}")
+                    return None
+                return await response.json()
+
+    async def _get_reranker_http_async(
+        self,
+        documents: List[str],
+        query: str,
+        top_n: Optional[int] = None,
+        model: str = None,
+        max_chunks_per_doc: Optional[int] = None,
+        return_documents: Optional[bool] = True,
+        return_len: Optional[bool] = True,
+        timeout: aiohttp.ClientTimeout = None, ** kwargs
+    ):
+        """对外暴露的reranker请求方法（包含重试）"""
+        return await self._retry_request(
+            self._get_reranker_http_core,
+            documents, query, top_n, model, max_chunks_per_doc,
+            return_documents, return_len, timeout, **kwargs
+        )
+
+    async def get_embeddings(
+        self,
+        corpus: List[str],
+        model: str = None,
+        timeout: Optional[Union[int, float]] = None
+    ):
+        """
+        获取语料库的嵌入向量，结果顺序与输入语料库顺序一致
+
+        Args:
+            corpus: 待生成嵌入向量的文本列表
+            model: 可选，指定使用的embedding模型名称，默认使用bge-large-zh-v1.5
+            timeout: 可选，超时时间（秒）：
+                     - 传int/float：表示总超时时间（秒）
+                     - 不传/None：使用默认永不超时配置
+        """
+        request_timeout = None
+        if timeout is not None:
+            if isinstance(timeout, (int, float)):
+                request_timeout = aiohttp.ClientTimeout(total=timeout)
+            else:
+                SYLogger.warning(
+                    f"Invalid timeout type: {type(timeout)}, must be int/float, use default timeout")
+
+        actual_model = model or self.default_embedding_model
+
+        SYLogger.info(
+            f"Requesting embeddings for corpus: {len(corpus)} items (model: {actual_model}, max_concurrency: {self.max_concurrency}, timeout: {timeout or 'None'})")
+
+        all_vectors = []
+
+        # [修复] 增加 Chunk 处理逻辑，防止 corpus 过大导致内存溢出或协程过多
+        # 每次最多处理 max_concurrency * 2 个请求，避免一次性创建几十万个协程
+        batch_size = self.max_concurrency * 2
+
+        for i in range(0, len(corpus), batch_size):
+            batch_texts = corpus[i: i + batch_size]
+
+            # 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]
+            results = await asyncio.gather(*tasks)
+
+            for result in results:
+                if result is None:
+                    # [修复] 尝试获取真实维度或使用配置兜底，不再硬编码 1024
+                    dim = self._model_dim_cache.get(actual_model)
+
+                    # 如果缓存中没有维度，尝试从配置对象获取（假设Config类有dimension属性）
+                    if dim is None:
+                        try:
+                            config = EmbeddingConfig.from_config(actual_model)
+                            if hasattr(config, 'dimension'):
+                                dim = config.dimension
+                            else:
+                                # 最后的兜底：如果配置也没有，必须有一个默认值防止崩溃
+                                # bge-large 通常是 1024
+                                dim = 1024
+                                SYLogger.warning(
+                                    f"Cannot get dimension from config for {actual_model}, use default 1024")
+                        except Exception:
+                            dim = 1024
+
+                    zero_vector = [0.0] * dim
+                    all_vectors.append(zero_vector)
+                    SYLogger.warning(
+                        f"Embedding request failed, append zero vector ({dim}D) for model {actual_model}")
+                    continue
+
+                # 从返回结果中提取向量并更新维度缓存
+                # 正常情况下 result["data"] 是一个列表
+                try:
+                    for item in result["data"]:
+                        embedding = item["embedding"]
+                        # [修复] 动态学习并缓存维度
+                        if actual_model not in self._model_dim_cache:
+                            self._model_dim_cache[actual_model] = len(
+                                embedding)
+                        all_vectors.append(embedding)
+                except (KeyError, TypeError) as e:
+                    SYLogger.error(f"Failed to parse embedding result: {e}")
+                    # 解析失败也补零
+                    dim = self._model_dim_cache.get(actual_model, 1024)
+                    all_vectors.append([0.0] * dim)
+
+        SYLogger.info(
+            f"Embeddings for corpus created: {len(all_vectors)} vectors (model: {actual_model})")
+        return all_vectors
+
+    async def get_reranker(
+        self,
+        top_results: List[str],
+        query: str,
+        model: str = None,
+        timeout: Optional[Union[int, float]] = None
+    ):
+        """
+        对搜索结果进行重排序
+
+        Args:
+            top_results: 待重排序的文本列表
+            query: 排序参考的查询语句
+            model: 可选，指定使用的reranker模型名称，默认使用bge-reranker-large
+            timeout: 可选，超时时间（秒）：
+                     - 传int/float：表示总超时时间（秒）
+                     - 不传/None：使用默认永不超时配置
+        """
+        request_timeout = None
+        if timeout is not None:
+            if isinstance(timeout, (int, float)):
+                request_timeout = aiohttp.ClientTimeout(total=timeout)
+            else:
+                SYLogger.warning(
+                    f"Invalid timeout type: {type(timeout)}, must be int/float, use default timeout")
+
+        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'})")
+
+        data = await self._get_reranker_http_async(
+            top_results, query, model=model, timeout=request_timeout)
+        SYLogger.info(
+            f"Reranker for top_results completed (model: {actual_model})")
+        return data

{sycommon_python_lib-0.1.57b2 → sycommon_python_lib-0.1.57b3}/src/sycommon_python_lib.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sycommon-python-lib
-Version: 0.1.57b2
+Version: 0.1.57b3
 Summary: Add your description here
 Requires-Python: >=3.11
 Description-Content-Type: text/markdown

sycommon_python_lib-0.1.57b2/src/sycommon/llm/embedding.py

@@ -1,204 +0,0 @@
-import asyncio
-import json
-import aiohttp
-from typing import Union, List, Optional
-
-from sycommon.config.Config import SingletonMeta
-from sycommon.config.EmbeddingConfig import EmbeddingConfig
-from sycommon.config.RerankerConfig import RerankerConfig
-from sycommon.logging.kafka_log import SYLogger
-
-
-class Embedding(metaclass=SingletonMeta):
-    def __init__(self):
-        # 1. 并发限制
-        self.max_concurrency = 20
-        # 保留默认模型名称
-        self.default_embedding_model = "bge-large-zh-v1.5"
-        self.default_reranker_model = "bge-reranker-large"
-
-        # 初始化默认模型的基础URL
-        self.embeddings_base_url = EmbeddingConfig.from_config(
-            self.default_embedding_model).baseUrl
-        self.reranker_base_url = RerankerConfig.from_config(
-            self.default_reranker_model).baseUrl
-
-        # 并发信号量
-        self.semaphore = asyncio.Semaphore(self.max_concurrency)
-        # 全局默认超时：永不超时（None）
-        self.default_timeout = aiohttp.ClientTimeout(total=None)
-
-    async def _get_embeddings_http_async(
-        self,
-        input: Union[str, List[str]],
-        encoding_format: str = None,
-        model: str = None,
-        timeout: aiohttp.ClientTimeout = None,
-        **kwargs
-    ):
-        async with self.semaphore:
-            # 优先使用传入的超时，无则用全局默认
-            request_timeout = timeout or self.default_timeout
-
-            # 优先使用传入的模型名，无则用默认值
-            target_model = model or self.default_embedding_model
-            target_base_url = EmbeddingConfig.from_config(target_model).baseUrl
-            url = f"{target_base_url}/v1/embeddings"
-
-            request_body = {
-                "model": target_model,
-                "input": input,
-                "encoding_format": encoding_format or "float"
-            }
-            request_body.update(kwargs)
-
-            try:
-                async with aiohttp.ClientSession(timeout=request_timeout) as session:
-                    async with session.post(url, json=request_body) as response:
-                        if response.status != 200:
-                            error_detail = await response.text()
-                            SYLogger.error(
-                                f"Embedding request failed (model: {target_model}): {error_detail}")
-                            return None
-                        return await response.json()
-            except asyncio.TimeoutError:
-                SYLogger.error(
-                    f"Embedding request timeout (model: {target_model})")
-                return None
-            except Exception as e:
-                SYLogger.error(
-                    f"Embedding request unexpected error (model: {target_model}): {str(e)}")
-                return None
-
-    async def _get_reranker_http_async(
-        self,
-        documents: List[str],
-        query: str,
-        top_n: Optional[int] = None,
-        model: str = None,
-        max_chunks_per_doc: Optional[int] = None,
-        return_documents: Optional[bool] = True,
-        return_len: Optional[bool] = True,
-        timeout: aiohttp.ClientTimeout = None,
-        **kwargs
-    ):
-        async with self.semaphore:
-            # 优先使用传入的超时，无则用全局默认
-            request_timeout = timeout or self.default_timeout
-
-            # 优先使用传入的模型名，无则用默认值
-            target_model = model or self.default_reranker_model
-            target_base_url = RerankerConfig.from_config(target_model).baseUrl
-            url = f"{target_base_url}/v1/rerank"
-
-            request_body = {
-                "model": target_model,
-                "documents": documents,
-                "query": query,
-                "top_n": top_n or len(documents),
-                "max_chunks_per_doc": max_chunks_per_doc,
-                "return_documents": return_documents,
-                "return_len": return_len,
-                "kwargs": json.dumps(kwargs),
-            }
-            request_body.update(kwargs)
-
-            try:
-                async with aiohttp.ClientSession(timeout=request_timeout) as session:
-                    async with session.post(url, json=request_body) as response:
-                        if response.status != 200:
-                            error_detail = await response.text()
-                            SYLogger.error(
-                                f"Rerank request failed (model: {target_model}): {error_detail}")
-                            return None
-                        return await response.json()
-            except asyncio.TimeoutError:
-                SYLogger.error(
-                    f"Rerank request timeout (model: {target_model})")
-                return None
-            except Exception as e:
-                SYLogger.error(
-                    f"Rerank request unexpected error (model: {target_model}): {str(e)}")
-                return None
-
-    async def get_embeddings(
-        self,
-        corpus: List[str],
-        model: str = None,
-        timeout: Optional[Union[int, float]] = None
-    ):
-        """
-        获取语料库的嵌入向量，结果顺序与输入语料库顺序一致
-
-        Args:
-            corpus: 待生成嵌入向量的文本列表
-            model: 可选，指定使用的embedding模型名称，默认使用bge-large-zh-v1.5
-            timeout: 可选，超时时间（秒）：
-                     - 传int/float：表示总超时时间（秒）
-                     - 不传/None：使用默认永不超时配置
-        """
-        request_timeout = None
-        if timeout is not None:
-            if isinstance(timeout, (int, float)):
-                request_timeout = aiohttp.ClientTimeout(total=timeout)
-            else:
-                SYLogger.warning(
-                    f"Invalid timeout type: {type(timeout)}, must be int/float, use default timeout")
-
-        SYLogger.info(
-            f"Requesting embeddings for corpus: {corpus} (model: {model or self.default_embedding_model}, max_concurrency: {self.max_concurrency}, 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 corpus]
-        results = await asyncio.gather(*tasks)
-
-        vectors = []
-        for result in results:
-            if result is None:
-                zero_vector = [0.0] * 1024
-                vectors.append(zero_vector)
-                SYLogger.warning(
-                    f"Embedding request failed, append zero vector (1024D)")
-                continue
-            for item in result["data"]:
-                vectors.append(item["embedding"])
-
-        SYLogger.info(
-            f"Embeddings for corpus: {corpus} created (model: {model or self.default_embedding_model})")
-        return vectors
-
-    async def get_reranker(
-        self,
-        top_results: List[str],
-        query: str,
-        model: str = None,
-        timeout: Optional[Union[int, float]] = None
-    ):
-        """
-        对搜索结果进行重排序
-
-        Args:
-            top_results: 待重排序的文本列表
-            query: 排序参考的查询语句
-            model: 可选，指定使用的reranker模型名称，默认使用bge-reranker-large
-            timeout: 可选，超时时间（秒）：
-                     - 传int/float：表示总超时时间（秒）
-                     - 不传/None：使用默认永不超时配置
-        """
-        request_timeout = None
-        if timeout is not None:
-            if isinstance(timeout, (int, float)):
-                request_timeout = aiohttp.ClientTimeout(total=timeout)
-            else:
-                SYLogger.warning(
-                    f"Invalid timeout type: {type(timeout)}, must be int/float, use default timeout")
-
-        SYLogger.info(
-            f"Requesting reranker for top_results: {top_results} (model: {model or self.default_reranker_model}, max_concurrency: {self.max_concurrency}, timeout: {timeout or 'None'})")
-
-        data = await self._get_reranker_http_async(
-            top_results, query, model=model, timeout=request_timeout)
-        SYLogger.info(
-            f"Reranker for top_results: {top_results} completed (model: {model or self.default_reranker_model})")
-        return data