feat(vertex): 集成 Vertex AI Express API 支持

本次更新引入了对 Google Vertex AI Express API 的支持，允许用户配置和使用 Vertex AI 模型。

主要变更包括：

后端：
- 新增 `VERTEX_API_KEYS` 和 `VERTEX_EXPRESS_BASE_URL` 至系统配置 ([`.env.example`](.env.example:13), [`app/config/config.py:62`](app/config/config.py:62), [`app/database/models.py`](app/database/models.py), [`app/database/services.py`](app/database/services.py))。
- 实现 `VertexExpressChatService` ([`app/service/chat/vertex_express_chat_service.py`](app/service/chat/vertex_express_chat_service.py)) 用于处理与 Vertex AI Express API 的交互。
- 添加 `vertex_express_routes` ([`app/router/vertex_express_routes.py`](app/router/vertex_express_routes.py)) 来暴露 Vertex AI 相关的 API 端点，并集成到主应用 ([`app/core/application.py:36`](app/core/application.py:36), [`app/router/routes.py:15`](app/router/routes.py:15))。
- 更新密钥管理器 ([`app/service/key/key_manager.py`](app/service/key/key_manager.py)) 以支持 Vertex API 密钥的获取、检查和轮换。

前端 (配置编辑器):
- 在配置页面 ([`app/templates/config_editor.html:463`](app/templates/config_editor.html:463)) 添加了 Vertex API 密钥列表和 Vertex Express API 基础 URL 的表单字段。
- 实现了批量添加和删除 Vertex API 密钥的功能，包括相应的模态框和操作逻辑 ([`app/static/js/config_editor.js:550`](app/static/js/config_editor.js:550), [`app/static/js/config_editor.js:1097`](app/static/js/config_editor.js:1097), [`app/templates/config_editor.html:1657`](app/templates/config_editor.html:1657))。
- 确保新的配置项在初始化 ([`app/static/js/config_editor.js:598`](app/static/js/config_editor.js:598)) 和表单填充 ([`app/static/js/config_editor.js:671`](app/static/js/config_editor.js:671)) 时得到正确处理。
- 更新了数组项添加逻辑以识别 `VERTEX_API_KEYS` 为敏感字段 ([`app/static/js/config_editor.js:1235`](app/static/js/config_editor.js:1235))。

此功能扩展了应用支持的 AI 服务范围，为用户提供了更多模型选择。

app/service/chat/gemini_chat_service.py

@@ -131,7 +131,7 @@ class GeminiChatService:
         self, original_response: Dict[str, Any], text: str
     ) -> Dict[str, Any]:
         """创建包含指定文本的响应"""
-        response_copy = json.loads(json.dumps(original_response))  # 深拷贝
+        response_copy = json.loads(json.dumps(original_response))
         if response_copy.get("candidates") and response_copy["candidates"][0].get(
             "content", {}
         ).get("parts"):
@@ -200,7 +200,7 @@ class GeminiChatService:
             request_datetime = datetime.datetime.now()
             start_time = time.perf_counter()
             current_attempt_key = api_key
-            final_api_key = current_attempt_key # Update final key used
+            final_api_key = current_attempt_key
             try:
                 async for line in self.api_client.stream_generate_content(
                     payload, model, current_attempt_key

app/service/chat/openai_chat_service.py

@@ -51,7 +51,7 @@ def _build_tools(
             or model.endswith("-image")
             or model.endswith("-image-generation")
         )
-        and not _has_media_parts(messages)  # Use the updated check
+        and not _has_media_parts(messages)
     ):
         tool["codeExecution"] = {}
         logger.debug("Code execution tool enabled.")

app/service/chat/vertex_express_chat_service.py

@@ -0,0 +1,277 @@
+# app/services/chat_service.py
+
+import json
+import re
+import datetime
+import time
+from typing import Any, AsyncGenerator, Dict, List
+from app.config.config import settings
+from app.core.constants import GEMINI_2_FLASH_EXP_SAFETY_SETTINGS
+from app.domain.gemini_models import GeminiRequest
+from app.handler.response_handler import GeminiResponseHandler
+from app.handler.stream_optimizer import gemini_optimizer
+from app.log.logger import get_gemini_logger
+from app.service.client.api_client import GeminiApiClient
+from app.service.key.key_manager import KeyManager
+from app.database.services import add_error_log, add_request_log
+
+logger = get_gemini_logger()
+
+
+def _has_image_parts(contents: List[Dict[str, Any]]) -> bool:
+    """判断消息是否包含图片部分"""
+    for content in contents:
+        if "parts" in content:
+            for part in content["parts"]:
+                if "image_url" in part or "inline_data" in part:
+                    return True
+    return False
+
+
+def _build_tools(model: str, payload: Dict[str, Any]) -> List[Dict[str, Any]]:
+    """构建工具"""
+    
+    def _merge_tools(tools: List[Dict[str, Any]]) -> Dict[str, Any]:
+        record = dict()
+        for item in tools:
+            if not item or not isinstance(item, dict):
+                continue
+
+            for k, v in item.items():
+                if k == "functionDeclarations" and v and isinstance(v, list):
+                    functions = record.get("functionDeclarations", [])
+                    functions.extend(v)
+                    record["functionDeclarations"] = functions
+                else:
+                    record[k] = v
+        return record
+
+    tool = dict()
+    if payload and isinstance(payload, dict) and "tools" in payload:
+        if payload.get("tools") and isinstance(payload.get("tools"), dict):
+            payload["tools"] = [payload.get("tools")]
+        items = payload.get("tools", [])
+        if items and isinstance(items, list):
+            tool.update(_merge_tools(items))
+
+    if (
+        settings.TOOLS_CODE_EXECUTION_ENABLED
+        and not (model.endswith("-search") or "-thinking" in model)
+        and not _has_image_parts(payload.get("contents", []))
+    ):
+        tool["codeExecution"] = {}
+    if model.endswith("-search"):
+        tool["googleSearch"] = {}
+
+    # 解决 "Tool use with function calling is unsupported" 问题
+    if tool.get("functionDeclarations"):
+        tool.pop("googleSearch", None)
+        tool.pop("codeExecution", None)
+
+    return [tool] if tool else []
+
+
+def _get_safety_settings(model: str) -> List[Dict[str, str]]:
+    """获取安全设置"""
+    if model == "gemini-2.0-flash-exp":
+        return GEMINI_2_FLASH_EXP_SAFETY_SETTINGS
+    return settings.SAFETY_SETTINGS
+
+
+def _build_payload(model: str, request: GeminiRequest) -> Dict[str, Any]:
+    """构建请求payload"""
+    request_dict = request.model_dump()
+    if request.generationConfig:
+        if request.generationConfig.maxOutputTokens is None:
+            # 如果未指定最大输出长度，则不传递该字段，解决截断的问题
+            request_dict["generationConfig"].pop("maxOutputTokens")
+    
+    payload = {
+        "contents": request_dict.get("contents", []),
+        "tools": _build_tools(model, request_dict),
+        "safetySettings": _get_safety_settings(model),
+        "generationConfig": request_dict.get("generationConfig"),
+        "systemInstruction": request_dict.get("systemInstruction"),
+    }
+
+    if model.endswith("-image") or model.endswith("-image-generation"):
+        payload.pop("systemInstruction")
+        payload["generationConfig"]["responseModalities"] = ["Text", "Image"]
+        
+    if model.endswith("-non-thinking"):
+        payload["generationConfig"]["thinkingConfig"] = {"thinkingBudget": 0} 
+    if model in settings.THINKING_BUDGET_MAP:
+        payload["generationConfig"]["thinkingConfig"] = {"thinkingBudget": settings.THINKING_BUDGET_MAP.get(model,1000)}
+
+    return payload
+
+
+class GeminiChatService:
+    """聊天服务"""
+
+    def __init__(self, base_url: str, key_manager: KeyManager):
+        self.api_client = GeminiApiClient(base_url, settings.TIME_OUT)
+        self.key_manager = key_manager
+        self.response_handler = GeminiResponseHandler()
+
+    def _extract_text_from_response(self, response: Dict[str, Any]) -> str:
+        """从响应中提取文本内容"""
+        if not response.get("candidates"):
+            return ""
+
+        candidate = response["candidates"][0]
+        content = candidate.get("content", {})
+        parts = content.get("parts", [])
+
+        if parts and "text" in parts[0]:
+            return parts[0].get("text", "")
+        return ""
+
+    def _create_char_response(
+        self, original_response: Dict[str, Any], text: str
+    ) -> Dict[str, Any]:
+        """创建包含指定文本的响应"""
+        response_copy = json.loads(json.dumps(original_response))  # 深拷贝
+        if response_copy.get("candidates") and response_copy["candidates"][0].get(
+            "content", {}
+        ).get("parts"):
+            response_copy["candidates"][0]["content"]["parts"][0]["text"] = text
+        return response_copy
+
+    async def generate_content(
+        self, model: str, request: GeminiRequest, api_key: str
+    ) -> Dict[str, Any]:
+        """生成内容"""
+        payload = _build_payload(model, request)
+        start_time = time.perf_counter()
+        request_datetime = datetime.datetime.now()
+        is_success = False
+        status_code = None
+        response = None
+
+        try:
+            response = await self.api_client.generate_content(payload, model, api_key)
+            is_success = True
+            status_code = 200
+            return self.response_handler.handle_response(response, model, stream=False)
+        except Exception as e:
+            is_success = False
+            error_log_msg = str(e)
+            logger.error(f"Normal API call failed with error: {error_log_msg}")
+            match = re.search(r"status code (\d+)", error_log_msg)
+            if match:
+                status_code = int(match.group(1))
+            else:
+                status_code = 500
+
+            await add_error_log(
+                gemini_key=api_key,
+                model_name=model,
+                error_type="gemini-chat-non-stream",
+                error_log=error_log_msg,
+                error_code=status_code,
+                request_msg=payload
+            )
+            raise e
+        finally:
+            end_time = time.perf_counter()
+            latency_ms = int((end_time - start_time) * 1000)
+            await add_request_log(
+                model_name=model,
+                api_key=api_key,
+                is_success=is_success,
+                status_code=status_code,
+                latency_ms=latency_ms,
+                request_time=request_datetime
+            )
+
+    async def stream_generate_content(
+        self, model: str, request: GeminiRequest, api_key: str
+    ) -> AsyncGenerator[str, None]:
+        """流式生成内容"""
+        retries = 0
+        max_retries = settings.MAX_RETRIES
+        payload = _build_payload(model, request)
+        is_success = False
+        status_code = None
+        final_api_key = api_key
+
+        while retries < max_retries:
+            request_datetime = datetime.datetime.now()
+            start_time = time.perf_counter()
+            current_attempt_key = api_key
+            final_api_key = current_attempt_key # Update final key used
+            try:
+                async for line in self.api_client.stream_generate_content(
+                    payload, model, current_attempt_key
+                ):
+                    # print(line)
+                    if line.startswith("data:"):
+                        line = line[6:]
+                        response_data = self.response_handler.handle_response(
+                            json.loads(line), model, stream=True
+                        )
+                        text = self._extract_text_from_response(response_data)
+                        # 如果有文本内容，且开启了流式输出优化器，则使用流式输出优化器处理
+                        if text and settings.STREAM_OPTIMIZER_ENABLED:
+                            # 使用流式输出优化器处理文本输出
+                            async for (
+                                optimized_chunk
+                            ) in gemini_optimizer.optimize_stream_output(
+                                text,
+                                lambda t: self._create_char_response(response_data, t),
+                                lambda c: "data: " + json.dumps(c) + "\n\n",
+                            ):
+                                yield optimized_chunk
+                        else:
+                            # 如果没有文本内容（如工具调用等），整块输出
+                            yield "data: " + json.dumps(response_data) + "\n\n"
+                logger.info("Streaming completed successfully")
+                is_success = True
+                status_code = 200
+                break
+            except Exception as e:
+                retries += 1
+                is_success = False
+                error_log_msg = str(e)
+                logger.warning(
+                    f"Streaming API call failed with error: {error_log_msg}. Attempt {retries} of {max_retries}"
+                )
+                match = re.search(r"status code (\d+)", error_log_msg)
+                if match:
+                    status_code = int(match.group(1))
+                else:
+                    status_code = 500
+
+                await add_error_log(
+                    gemini_key=current_attempt_key,
+                    model_name=model,
+                    error_type="gemini-chat-stream",
+                    error_log=error_log_msg,
+                    error_code=status_code,
+                    request_msg=payload
+                )
+
+                api_key = await self.key_manager.handle_api_failure(current_attempt_key, retries)
+                if api_key:
+                    logger.info(f"Switched to new API key: {api_key}")
+                else:
+                    logger.error(f"No valid API key available after {retries} retries.")
+                    break
+
+                if retries >= max_retries:
+                    logger.error(
+                        f"Max retries ({max_retries}) reached for streaming."
+                    )
+                    break
+            finally:
+                end_time = time.perf_counter()
+                latency_ms = int((end_time - start_time) * 1000)
+                await add_request_log(
+                    model_name=model,
+                    api_key=final_api_key,
+                    is_success=is_success,
+                    status_code=status_code,
+                    latency_ms=latency_ms,
+                    request_time=request_datetime
+                )

app/service/config/config_service.py

@@ -55,7 +55,7 @@ class ConfigService:
             # 处理不同类型的值
             if isinstance(value, list):
                 db_value = json.dumps(value)
-            elif isinstance(value, dict):  # 新增对 dict 类型的处理
+            elif isinstance(value, dict):
                 db_value = json.dumps(value)
             elif isinstance(value, bool):
                 db_value = str(value).lower()
@@ -115,7 +115,7 @@ class ConfigService:
         # 重置并重新初始化 KeyManager
         try:
             await reset_key_manager_instance()
-            await get_key_manager_instance(settings.API_KEYS)
+            await get_key_manager_instance(settings.API_KEYS, settings.VERTEX_API_KEYS)
             logger.info("KeyManager instance re-initialized with updated settings.")
         except Exception as e:
             logger.error(f"Failed to re-initialize KeyManager: {str(e)}")
@@ -154,7 +154,7 @@ class ConfigService:
         deleted_count = 0
         not_found_keys: List[str] = []
 
-        current_api_keys = list(settings.API_KEYS)  # 创建副本以进行修改
+        current_api_keys = list(settings.API_KEYS)
         keys_actually_removed: List[str] = []
 
         for key_to_del in keys_to_delete:
@@ -166,7 +166,7 @@ class ConfigService:
                 not_found_keys.append(key_to_del)
 
         if deleted_count > 0:
-            settings.API_KEYS = current_api_keys  # 更新内存中的 settings
+            settings.API_KEYS = current_api_keys
             await ConfigService.update_config({"API_KEYS": settings.API_KEYS})
             logger.info(
                 f"成功删除 {deleted_count} 个密钥。密钥: {keys_actually_removed}"
@@ -182,9 +182,9 @@ class ConfigService:
             }
         else:
             message = "没有密钥被删除。"
-            if not_found_keys:  # 如果提供了密钥但都未找到
+            if not_found_keys:
                 message = f"所有 {len(not_found_keys)} 个指定的密钥均未找到: {not_found_keys}。"
-            elif not keys_to_delete:  # 如果 keys_to_delete 列表为空
+            elif not keys_to_delete:
                 message = "未指定要删除的密钥。"
             logger.warning(message)
             return {

app/service/error_log/error_log_service.py

@@ -161,10 +161,9 @@ async def process_delete_all_error_logs() -> int:
     返回删除的日志数量。
     """
     try:
-        # 确保数据库已连接 (如果适用，类似于 delete_old_error_logs)
-        # if not database.is_connected:
-        #     await database.connect()
-        #     logger.info("Database connection established for deleting all error logs.")
+        if not database.is_connected:
+            await database.connect()
+            logger.info("Database connection established for deleting all error logs.")
 
         deleted_count = await db_services.delete_all_error_logs()
         logger.info(

app/service/key/key_manager.py

@@ -9,12 +9,19 @@ logger = get_key_manager_logger()
 
 
 class KeyManager:
-    def __init__(self, api_keys: list):
+    def __init__(self, api_keys: list, vertex_api_keys: list):
         self.api_keys = api_keys
+        self.vertex_api_keys = vertex_api_keys
         self.key_cycle = cycle(api_keys)
+        self.vertex_key_cycle = cycle(vertex_api_keys)
         self.key_cycle_lock = asyncio.Lock()
+        self.vertex_key_cycle_lock = asyncio.Lock()
         self.failure_count_lock = asyncio.Lock()
+        self.vertex_failure_count_lock = asyncio.Lock()
         self.key_failure_counts: Dict[str, int] = {key: 0 for key in api_keys}
+        self.vertex_key_failure_counts: Dict[str, int] = {
+            key: 0 for key in vertex_api_keys
+        }
         self.MAX_FAILURES = settings.MAX_FAILURES
         self.paid_key = settings.PAID_KEY
 
@@ -26,17 +33,33 @@ class KeyManager:
         async with self.key_cycle_lock:
             return next(self.key_cycle)
 
+    async def get_next_vertex_key(self) -> str:
+        """获取下一个 Vertex API key"""
+        async with self.vertex_key_cycle_lock:
+            return next(self.vertex_key_cycle)
+
     async def is_key_valid(self, key: str) -> bool:
         """检查key是否有效"""
         async with self.failure_count_lock:
             return self.key_failure_counts[key] < self.MAX_FAILURES
 
+    async def is_vertex_key_valid(self, key: str) -> bool:
+        """检查 Vertex key 是否有效"""
+        async with self.vertex_failure_count_lock:
+            return self.vertex_key_failure_counts[key] < self.MAX_FAILURES
+
     async def reset_failure_counts(self):
         """重置所有key的失败计数"""
         async with self.failure_count_lock:
             for key in self.key_failure_counts:
                 self.key_failure_counts[key] = 0
 
+    async def reset_vertex_failure_counts(self):
+        """重置所有 Vertex key 的失败计数"""
+        async with self.vertex_failure_count_lock:
+            for key in self.vertex_key_failure_counts:
+                self.vertex_key_failure_counts[key] = 0
+
     async def reset_key_failure_count(self, key: str) -> bool:
         """重置指定key的失败计数"""
         async with self.failure_count_lock:
@@ -49,6 +72,18 @@ class KeyManager:
             )
             return False
 
+    async def reset_vertex_key_failure_count(self, key: str) -> bool:
+        """重置指定 Vertex key 的失败计数"""
+        async with self.vertex_failure_count_lock:
+            if key in self.vertex_key_failure_counts:
+                self.vertex_key_failure_counts[key] = 0
+                logger.info(f"Reset failure count for Vertex key: {key}")
+                return True
+            logger.warning(
+                f"Attempt to reset failure count for non-existent Vertex key: {key}"
+            )
+            return False
+
     async def get_next_working_key(self) -> str:
         """获取下一可用的API key"""
         initial_key = await self.get_next_key()
@@ -62,6 +97,19 @@ class KeyManager:
             if current_key == initial_key:
                 return current_key
 
+    async def get_next_working_vertex_key(self) -> str:
+        """获取下一可用的 Vertex API key"""
+        initial_key = await self.get_next_vertex_key()
+        current_key = initial_key
+
+        while True:
+            if await self.is_vertex_key_valid(current_key):
+                return current_key
+
+            current_key = await self.get_next_vertex_key()
+            if current_key == initial_key:
+                return current_key
+
     async def handle_api_failure(self, api_key: str, retries: int) -> str:
         """处理API调用失败"""
         async with self.failure_count_lock:
@@ -75,10 +123,23 @@ class KeyManager:
         else:
             return ""
 
+    async def handle_vertex_api_failure(self, api_key: str, retries: int) -> str:
+        """处理 Vertex API 调用失败"""
+        async with self.vertex_failure_count_lock:
+            self.vertex_key_failure_counts[api_key] += 1
+            if self.vertex_key_failure_counts[api_key] >= self.MAX_FAILURES:
+                logger.warning(
+                    f"Vertex API key {api_key} has failed {self.MAX_FAILURES} times"
+                )
+
     def get_fail_count(self, key: str) -> int:
         """获取指定密钥的失败次数"""
         return self.key_failure_counts.get(key, 0)
 
+    def get_vertex_fail_count(self, key: str) -> int:
+        """获取指定 Vertex 密钥的失败次数"""
+        return self.vertex_key_failure_counts.get(key, 0)
+
     async def get_keys_by_status(self) -> dict:
         """获取分类后的API key列表，包括失败次数"""
         valid_keys = {}
@@ -94,6 +155,20 @@ class KeyManager:
 
         return {"valid_keys": valid_keys, "invalid_keys": invalid_keys}
 
+    async def get_vertex_keys_by_status(self) -> dict:
+        """获取分类后的 Vertex API key 列表，包括失败次数"""
+        valid_keys = {}
+        invalid_keys = {}
+
+        async with self.vertex_failure_count_lock:
+            for key in self.vertex_api_keys:
+                fail_count = self.vertex_key_failure_counts[key]
+                if fail_count < self.MAX_FAILURES:
+                    valid_keys[key] = fail_count
+                else:
+                    invalid_keys[key] = fail_count
+        return {"valid_keys": valid_keys, "invalid_keys": invalid_keys}
+
     async def get_first_valid_key(self) -> str:
         """获取第一个有效的API key"""
         async with self.failure_count_lock:
@@ -111,19 +186,24 @@ class KeyManager:
 _singleton_instance = None
 _singleton_lock = asyncio.Lock()
 _preserved_failure_counts: Dict[str, int] | None = None
+_preserved_vertex_failure_counts: Dict[str, int] | None = None
 _preserved_old_api_keys_for_reset: list | None = None
+_preserved_vertex_old_api_keys_for_reset: list | None = None
 _preserved_next_key_in_cycle: str | None = None
+_preserved_vertex_next_key_in_cycle: str | None = None
 
 
-async def get_key_manager_instance(api_keys: list = None) -> KeyManager:
+async def get_key_manager_instance(
+    api_keys: list = None, vertex_api_keys: list = None
+) -> KeyManager:
     """
     获取 KeyManager 单例实例。
 
-    如果尚未创建实例，将使用提供的 api_keys 初始化 KeyManager。
+    如果尚未创建实例，将使用提供的 api_keys,vertex_api_keys 初始化 KeyManager。
     如果已创建实例，则忽略 api_keys 参数，返回现有单例。
     如果在重置后调用，会尝试恢复之前的状态（失败计数、循环位置）。
     """
-    global _singleton_instance, _preserved_failure_counts, _preserved_old_api_keys_for_reset, _preserved_next_key_in_cycle
+    global _singleton_instance, _preserved_failure_counts, _preserved_vertex_failure_counts, _preserved_old_api_keys_for_reset, _preserved_vertex_old_api_keys_for_reset, _preserved_next_key_in_cycle, _preserved_vertex_next_key_in_cycle
 
     async with _singleton_lock:
         if _singleton_instance is None:
@@ -131,14 +211,23 @@ async def get_key_manager_instance(api_keys: list = None) -> KeyManager:
                 raise ValueError(
                     "API keys are required to initialize or re-initialize the KeyManager instance."
                 )
+            if vertex_api_keys is None:
+                raise ValueError(
+                    "Vertex API keys are required to initialize or re-initialize the KeyManager instance."
+                )
+
             if not api_keys:
                 logger.warning(
                     "Initializing KeyManager with an empty list of API keys."
                 )
+            if not vertex_api_keys:
+                logger.warning(
+                    "Initializing KeyManager with an empty list of Vertex API keys."
+                )
 
-            _singleton_instance = KeyManager(api_keys)
+            _singleton_instance = KeyManager(api_keys, vertex_api_keys)
             logger.info(
-                f"KeyManager instance created/re-created with {len(api_keys)} API keys."
+                f"KeyManager instance created/re-created with {len(api_keys)} API keys and {len(vertex_api_keys)} Vertex API keys."
             )
 
             # 1. 恢复失败计数
@@ -153,6 +242,19 @@ async def get_key_manager_instance(api_keys: list = None) -> KeyManager:
                 logger.info("Inherited failure counts for applicable keys.")
             _preserved_failure_counts = None
 
+            if _preserved_vertex_failure_counts:
+                current_vertex_failure_counts = {
+                    key: 0 for key in _singleton_instance.vertex_api_keys
+                }
+                for key, count in _preserved_vertex_failure_counts.items():
+                    if key in current_vertex_failure_counts:
+                        current_vertex_failure_counts[key] = count
+                _singleton_instance.vertex_key_failure_counts = (
+                    current_vertex_failure_counts
+                )
+                logger.info("Inherited failure counts for applicable Vertex keys.")
+            _preserved_vertex_failure_counts = None
+
             # 2. 调整 key_cycle 的起始点
             start_key_for_new_cycle = None
             if (
@@ -201,9 +303,7 @@ async def get_key_manager_instance(api_keys: list = None) -> KeyManager:
                         f"Determined start key '{start_key_for_new_cycle}' not found in new API keys during cycle advancement. "
                         "New cycle will start from the beginning."
                     )
-                except (
-                    StopIteration
-                ):
+                except StopIteration:
                     logger.error(
                         "StopIteration while advancing key cycle, implies empty new API key list previously missed."
                     )
@@ -225,6 +325,76 @@ async def get_key_manager_instance(api_keys: list = None) -> KeyManager:
             _preserved_old_api_keys_for_reset = None
             _preserved_next_key_in_cycle = None
 
+            # 3. 调整 vertex_key_cycle 的起始点
+            start_key_for_new_vertex_cycle = None
+            if (
+                _preserved_vertex_old_api_keys_for_reset
+                and _preserved_vertex_next_key_in_cycle
+                and _singleton_instance.vertex_api_keys
+            ):
+                try:
+                    start_idx_in_old = _preserved_vertex_old_api_keys_for_reset.index(
+                        _preserved_vertex_next_key_in_cycle
+                    )
+
+                    for i in range(len(_preserved_vertex_old_api_keys_for_reset)):
+                        current_old_key_idx = (start_idx_in_old + i) % len(
+                            _preserved_vertex_old_api_keys_for_reset
+                        )
+                        key_candidate = _preserved_vertex_old_api_keys_for_reset[
+                            current_old_key_idx
+                        ]
+                        if key_candidate in _singleton_instance.vertex_api_keys:
+                            start_key_for_new_vertex_cycle = key_candidate
+                            break
+                except ValueError:
+                    logger.warning(
+                        f"Preserved next key '{_preserved_vertex_next_key_in_cycle}' not found in preserved old Vertex API keys. "
+                        "New cycle will start from the beginning of the new list."
+                    )
+                except Exception as e:
+                    logger.error(
+                        f"Error determining start key for new Vertex key cycle from preserved state: {e}. "
+                        "New cycle will start from the beginning."
+                    )
+                
+            if start_key_for_new_vertex_cycle and _singleton_instance.vertex_api_keys:
+                try:
+                    target_idx = _singleton_instance.vertex_api_keys.index(
+                        start_key_for_new_vertex_cycle
+                    )
+                    for _ in range(target_idx):
+                        next(_singleton_instance.vertex_key_cycle)
+                    logger.info(
+                        f"Vertex key cycle in new instance advanced. Next call to get_next_vertex_key() will yield: {start_key_for_new_vertex_cycle}"
+                    )
+                except ValueError:
+                    logger.warning(
+                        f"Determined start key '{start_key_for_new_vertex_cycle}' not found in new Vertex API keys during cycle advancement. "
+                        "New cycle will start from the beginning."
+                    )
+                except StopIteration:
+                    logger.error(
+                        "StopIteration while advancing Vertex key cycle, implies empty new Vertex API key list previously missed."
+                    )
+                except Exception as e:
+                    logger.error(
+                        f"Error advancing new Vertex key cycle: {e}. Cycle will start from beginning."
+                    ) 
+            else:
+                if _singleton_instance.vertex_api_keys:
+                    logger.info(
+                        "New Vertex key cycle will start from the beginning of the new Vertex API key list (no specific start key determined or needed)."
+                    )
+                else:
+                    logger.info(
+                        "New Vertex key cycle not applicable as the new Vertex API key list is empty."
+                    )
+
+            # 清理所有保存的状态
+            _preserved_vertex_old_api_keys_for_reset = None
+            _preserved_vertex_next_key_in_cycle = None
+
         return _singleton_instance
 
 
@@ -234,34 +404,52 @@ async def reset_key_manager_instance():
     将保存当前实例的状态（失败计数、旧 API keys、下一个 key 提示）
     以供下一次 get_key_manager_instance 调用时恢复。
     """
-    global _singleton_instance, _preserved_failure_counts, _preserved_old_api_keys_for_reset, _preserved_next_key_in_cycle
+    global _singleton_instance, _preserved_failure_counts, _preserved_vertex_failure_counts, _preserved_old_api_keys_for_reset, _preserved_vertex_old_api_keys_for_reset, _preserved_next_key_in_cycle, _preserved_vertex_next_key_in_cycle
     async with _singleton_lock:
         if _singleton_instance:
             # 1. 保存失败计数
             _preserved_failure_counts = _singleton_instance.key_failure_counts.copy()
+            _preserved_vertex_failure_counts = _singleton_instance.vertex_key_failure_counts.copy()
 
             # 2. 保存旧的 API keys 列表
             _preserved_old_api_keys_for_reset = _singleton_instance.api_keys.copy()
+            _preserved_vertex_old_api_keys_for_reset = _singleton_instance.vertex_api_keys.copy()
 
             # 3. 保存 key_cycle 的下一个 key 提示
             try:
                 if _singleton_instance.api_keys:
-                    _preserved_next_key_in_cycle = (
+                    _preserved_next_key_in_cycle = (    
                         await _singleton_instance.get_next_key()
                     )
                 else:
                     _preserved_next_key_in_cycle = None
-            except (
-                StopIteration
-            ):
+            except StopIteration:
                 logger.warning(
                     "Could not preserve next key hint: key cycle was empty or exhausted in old instance."
                 )
-                _preserved_next_key_in_cycle = None
+                _preserved_next_key_in_cycle = None 
             except Exception as e:
                 logger.error(f"Error preserving next key hint during reset: {e}")
                 _preserved_next_key_in_cycle = None
 
+            # 4. 保存 vertex_key_cycle 的下一个 key 提示
+            try:
+                if _singleton_instance.vertex_api_keys:
+                    _preserved_vertex_next_key_in_cycle = (
+                        await _singleton_instance.get_next_vertex_key()
+                    )
+                else:
+                    _preserved_vertex_next_key_in_cycle = None
+            except StopIteration:
+                logger.warning(
+                    "Could not preserve next key hint: Vertex key cycle was empty or exhausted in old instance."
+                )   
+                _preserved_vertex_next_key_in_cycle = None
+            except Exception as e:
+                logger.error(f"Error preserving next key hint during reset: {e}")
+                _preserved_vertex_next_key_in_cycle = None
+                
+
             _singleton_instance = None
             logger.info(
                 "KeyManager instance has been reset. State (failure counts, old keys, next key hint) preserved for next instantiation."

app/service/openai_compatiable/openai_compatiable_service.py

@@ -131,7 +131,7 @@ class OpenAICompatiableService:
                 logger.info("Streaming completed successfully")
                 is_success = True
                 status_code = 200
-                break  # 成功后退出循环
+                break
             except Exception as e:
                 retries += 1
                 is_success = False