mirror of https://github.com/Mai-with-u/MaiBot.git
Merge pull request #1251 from foxplaying/patch-2
Gemini：正式启用HttpOptions对传入参数的处理，优化预算提示，修正文档pull/1258/head^2
UnCLAS-Prommer
GitHub
commit
15d6a319d3
|
|
@ -28,7 +28,7 @@ version = "1.1.1"
|
|||
```toml
|
||||
[[api_providers]]
|
||||
name = "DeepSeek" # 服务商名称（自定义）
|
||||
base_url = "https://api.deepseek.cn/v1" # API服务的基础URL
|
||||
base_url = "https://api.deepseek.com/v1" # API服务的基础URL
|
||||
api_key = "your-api-key-here" # API密钥
|
||||
client_type = "openai" # 客户端类型
|
||||
max_retry = 2 # 最大重试次数
|
||||
|
|
@ -43,19 +43,19 @@ retry_interval = 10 # 重试间隔（秒）
|
|||
| `name` | ✅ | 服务商名称，需要在模型配置中引用 | - |
|
||||
| `base_url` | ✅ | API服务的基础URL | - |
|
||||
| `api_key` | ✅ | API密钥，请替换为实际密钥 | - |
|
||||
| `client_type` | ❌ | 客户端类型：`openai`（OpenAI格式）或 `gemini`（Gemini格式，现在支持不良好） | `openai` |
|
||||
| `client_type` | ❌ | 客户端类型：`openai`（OpenAI格式）或 `gemini`（Gemini格式） | `openai` |
|
||||
| `max_retry` | ❌ | API调用失败时的最大重试次数 | 2 |
|
||||
| `timeout` | ❌ | API请求超时时间（秒） | 30 |
|
||||
| `retry_interval` | ❌ | 重试间隔时间（秒） | 10 |
|
||||
|
||||
**请注意，对于`client_type`为`gemini`的模型，`base_url`字段无效。**
|
||||
**请注意，对于`client_type`为`gemini`的模型，`retry`字段由`gemini`自己决定。**
|
||||
### 2.3 支持的服务商示例
|
||||
|
||||
#### DeepSeek
|
||||
```toml
|
||||
[[api_providers]]
|
||||
name = "DeepSeek"
|
||||
base_url = "https://api.deepseek.cn/v1"
|
||||
base_url = "https://api.deepseek.com/v1"
|
||||
api_key = "your-deepseek-api-key"
|
||||
client_type = "openai"
|
||||
```
|
||||
|
|
@ -73,7 +73,7 @@ client_type = "openai"
|
|||
```toml
|
||||
[[api_providers]]
|
||||
name = "Google"
|
||||
base_url = "https://api.google.com/v1"
|
||||
base_url = "https://generativelanguage.googleapis.com/v1beta"
|
||||
api_key = "your-google-api-key"
|
||||
client_type = "gemini" # 注意：Gemini需要使用特殊客户端
|
||||
```
|
||||
|
|
@ -131,9 +131,20 @@ enable_thinking = false # 禁用思考
|
|||
[models.extra_params]
|
||||
thinking = {type = "disabled"} # 禁用思考
|
||||
```
|
||||
|
||||
而对于`gemini`需要单独进行配置
|
||||
```toml
|
||||
[[models]]
|
||||
model_identifier = "gemini-2.5-flash"
|
||||
name = "gemini-2.5-flash"
|
||||
api_provider = "Google"
|
||||
[models.extra_params]
|
||||
thinking_budget = 0 # 禁用思考
|
||||
# thinking_budget = -1 由模型自己决定
|
||||
```
|
||||
|
||||
请注意，`extra_params` 的配置应该构成一个合法的TOML字典结构，具体内容取决于API服务商的要求。
|
||||
|
||||
**请注意，对于`client_type`为`gemini`的模型，此字段无效。**
|
||||
### 3.3 配置参数说明
|
||||
|
||||
| 参数 | 必填 | 说明 |
|
||||
|
|
|
|||
|
|
@ -17,6 +17,7 @@ from google.genai.types import (
|
|||
EmbedContentResponse,
|
||||
EmbedContentConfig,
|
||||
SafetySetting,
|
||||
HttpOptions,
|
||||
HarmCategory,
|
||||
HarmBlockThreshold,
|
||||
)
|
||||
|
|
@ -345,22 +346,27 @@ class GeminiClient(BaseClient):
|
|||
|
||||
def __init__(self, api_provider: APIProvider):
|
||||
super().__init__(api_provider)
|
||||
|
||||
# 增加传入参数处理
|
||||
http_options_kwargs = {"timeout": api_provider.timeout}
|
||||
|
||||
# 秒转换为毫秒传入
|
||||
if api_provider.timeout is not None:
|
||||
http_options_kwargs["timeout"] = int(api_provider.timeout * 1000)
|
||||
|
||||
# 传入并处理地址和版本(必须为Gemini格式)
|
||||
if api_provider.base_url:
|
||||
parts = api_provider.base_url.rstrip("/").rsplit("/", 1)
|
||||
if len(parts) == 2 and parts[1].startswith("v"):
|
||||
http_options_kwargs["base_url"] = f"{parts[0]}/"
|
||||
http_options_kwargs["api_version"] = parts[1]
|
||||
else:
|
||||
http_options_kwargs["base_url"] = api_provider.base_url
|
||||
self.client = genai.Client(
|
||||
http_options=HttpOptions(**http_options_kwargs),
|
||||
api_key=api_provider.api_key,
|
||||
) # 这里和openai不一样，gemini会自己决定自己是否需要retry
|
||||
|
||||
# 尝试传入自定义base_url(实验性，必须为Gemini格式)
|
||||
if hasattr(api_provider, "base_url") and api_provider.base_url:
|
||||
base_url = api_provider.base_url.rstrip("/") # 去掉末尾 /
|
||||
self.client._api_client._http_options.base_url = base_url
|
||||
|
||||
# 如果 base_url 已经带了 /v1 或 /v1beta，就清掉 SDK 的 api_version
|
||||
if base_url.endswith("/v1") or base_url.endswith("/v1beta"):
|
||||
self.client._api_client._http_options.api_version = None
|
||||
|
||||
# 让 GeminiClient 内部也能访问底层 api_client
|
||||
self._api_client = self.client._api_client
|
||||
|
||||
@staticmethod
|
||||
def clamp_thinking_budget(tb: int, model_id: str) -> int:
|
||||
"""
|
||||
|
|
@ -380,20 +386,29 @@ class GeminiClient(BaseClient):
|
|||
limits = THINKING_BUDGET_LIMITS[key]
|
||||
break
|
||||
|
||||
# 特殊值处理
|
||||
# 预算值处理
|
||||
if tb == THINKING_BUDGET_AUTO:
|
||||
return THINKING_BUDGET_AUTO
|
||||
if tb == THINKING_BUDGET_DISABLED:
|
||||
if limits and limits.get("can_disable", False):
|
||||
return THINKING_BUDGET_DISABLED
|
||||
return limits["min"] if limits else THINKING_BUDGET_AUTO
|
||||
if limits:
|
||||
logger.warning(f"模型 {model_id} 不支持禁用思考预算，已回退到最小值 {limits['min']}")
|
||||
return limits["min"]
|
||||
return THINKING_BUDGET_AUTO
|
||||
|
||||
# 已知模型裁剪到范围
|
||||
# 已知模型范围裁剪 + 提示
|
||||
if limits:
|
||||
return max(limits["min"], min(tb, limits["max"]))
|
||||
if tb < limits["min"]:
|
||||
logger.warning(f"模型 {model_id} 的 thinking_budget={tb} 过小，已调整为最小值 {limits['min']}")
|
||||
return limits["min"]
|
||||
if tb > limits["max"]:
|
||||
logger.warning(f"模型 {model_id} 的 thinking_budget={tb} 过大，已调整为最大值 {limits['max']}")
|
||||
return limits["max"]
|
||||
return tb
|
||||
|
||||
# 未知模型，返回动态模式
|
||||
logger.warning(f"模型 {model_id} 未在 THINKING_BUDGET_LIMITS 中定义，将使用动态模式 tb=-1 兼容。")
|
||||
# 未知模型 → 默认自动模式
|
||||
logger.warning(f"模型 {model_id} 未在 THINKING_BUDGET_LIMITS 中定义，已启用模型自动预算兼容")
|
||||
return THINKING_BUDGET_AUTO
|
||||
|
||||
async def get_response(
|
||||
|
|
@ -448,7 +463,7 @@ class GeminiClient(BaseClient):
|
|||
try:
|
||||
tb = int(extra_params["thinking_budget"])
|
||||
except (ValueError, TypeError):
|
||||
logger.warning(f"无效的 thinking_budget 值 {extra_params['thinking_budget']}，将使用默认动态模式 {tb}")
|
||||
logger.warning(f"无效的 thinking_budget 值 {extra_params['thinking_budget']}，将使用模型自动预算模式 {tb}")
|
||||
# 裁剪到模型支持的范围
|
||||
tb = self.clamp_thinking_budget(tb, model_info.model_identifier)
|
||||
|
||||
|
|
|
|||
Loading…
Reference in New Issue