Skip to content

1.添加了阿里千问模型，由于其不支持pydantic，补充了一些适配 #254

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
hairuchen wants to merge 3 commits into
base: main
Choose a base branch
from
Open

1.添加了阿里千问模型，由于其不支持pydantic，补充了一些适配 #254

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
089ac09
1.添加了阿里千问模型，由于其不支持pydantic，补充了一些适配
hairuchen Dec 22, 2025
6ac0808
代码修改
hairuchen Dec 22, 2025
036ed18
代码修改
hairuchen Dec 22, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 5 additions & 0 deletions stagehand/browser.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -180,8 +180,11 @@ async def connect_local_browser(
except Exception as e:
logger.error(f"Failed to create downloads_path {downloads_path}: {e}")

executable_path_option = local_browser_launch_options.get("executablePath")

# Prepare Launch Options (translate keys if needed)
launch_options = {
"executable_path": executable_path_option,
"headless": local_browser_launch_options.get(
"headless", stagehand_instance.config.headless
),
Expand All @@ -208,6 +211,8 @@ async def connect_local_browser(
"ignoreHTTPSErrors", True
),
}
if executable_path_option:
launch_options["executable_path"] = executable_path_option
launch_options = {k: v for k, v in launch_options.items() if v is not None}

# Launch Context
Expand Down
194 changes: 194 additions & 0 deletions stagehand/llm/qwenclient.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,194 @@
from typing import Any, Dict, Optional

import json
import aiohttp
from pydantic import BaseModel
from stagehand.llm.client import LLMClient # 继承项目现有的 LLMClient 基类
from stagehand.metrics import start_inference_timer, get_inference_time_ms


class HybridDict(dict):
def __init__(self, data: dict):
super().__init__(data)
# 递归处理嵌套字典（如 usage、choices 里的内容）
for k, v in data.items():
if isinstance(v, dict):
self[k] = HybridDict(v)
elif isinstance(v, list):
# 处理列表中的字典（如 choices 数组）
self[k] = [HybridDict(item) if isinstance(item, dict) else item for item in v]

# 支持属性访问（如 self.usage → self['usage']）
def __getattr__(self, name):
try:
return self[name]
except KeyError:
raise AttributeError(f"'HybridDict' object has no attribute '{name}'")

# 支持属性赋值（可选）
def __setattr__(self, name, value):
self[name] = value


class QwenClient(LLMClient):
def __init__(self, stagehand_logger, api_key: str, model_name: str = "qwen-turbo", **kwargs):
# 调用父类构造函数，符合现有 LLMClient 的初始化方式
super().__init__(
stagehand_logger=stagehand_logger,
api_key=api_key,
default_model=model_name, **kwargs
)
self.api_key = api_key
self.api_base = "https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions"

async def create_response(
self,
*,
messages: list[dict[str, str]],
model: Optional[str] = None,
function_name: Optional[str] = None,
**kwargs: Any,
) -> dict[str, Any]:
# 1. 基础参数校验
model = model or self.default_model
if not model:
raise ValueError("未指定模型名称")

# 2. 构建请求头（修复：自定义 headers，而非引用 litellm 模块）
headers = {
"Content-Type": "application/json",
"Authorization": f"Bearer {self.model_api_key}"
}

# 3. 处理 response_format + 自动追加 JSON 格式提示
response_format = kwargs.get("response_format")
# 深拷贝 messages，避免修改原始数据（关键！）
processed_messages = [msg.copy() for msg in messages]

# 仅当指定 response_format 时，追加 JSON 格式要求
if response_format:
json_format_prompt = """
你是浏览器自动化的元素识别助手，需基于 Accessibility Tree（语义化节点树）返回符合要求的可操作元素，规则如下：
### 输入说明
1. Accessibility Tree 包含节点格式：[节点ID] 角色: 标签（如 [4] textbox: 请输入账号）；
2. 目标操作是用户的自然语言指令（如「在请输入账号输入框中输入内容」）。
### 返回规则
1. 仅返回合法 JSON 字符串，无任何多余文字/解释/代码块；
2. 顶层为字典，仅包含 "elements" 键（值为数组）；
3. 数组内每个元素必须包含以下字段：
- element_id：Accessibility Tree 中的原始数字 ID（如 [4] 中的 4，整数类型，用于定位元素）；
- description：元素的描述（结合角色和标签，如「textbox: 请输入账号」）；
- method：Playwright 支持的操作方法（如 textbox/textarea 用 "fill"，button/link 用 "click"）；
- arguments：必须是list数组结构，操作的参数列表list,即使只有一个元素也应该用数组嵌套，没有元素时填入一个空字符串''。
### 返回格式示例
{
"elements": [
{
"element_id": 4,
"description": "textbox: 请输入账号",
"method": "fill",
"arguments": ["15211228071"]
}
]
}
### 强制要求
1. 仅返回上述格式的 JSON，无其他内容；
2. element_id 必须是 Accessibility Tree 中的原始数字，不允许自定义；
3. 若未找到匹配元素，返回 {"elements": []}；
4. 若用户指令是动作（如输入/点击），优先返回最匹配的单个元素；若为观察（如"找到所有按钮"），返回所有符合条件的元素。
"""
has_system_msg = False

# 遍历 messages，在已有 system 消息后追加提示
for msg in processed_messages:
if msg["role"] == "system":
msg["content"] += json_format_prompt
has_system_msg = True
break

# 若无 system 消息，新增一条（保证 JSON 提示存在）
if not has_system_msg:
processed_messages.insert(0, {
"role": "system",
"content": f"你是一个专业的助手{json_format_prompt}"
})

# 4. 初始化请求体（使用处理后的 messages）
payload = {
"model": model,
"messages": processed_messages, # 用追加提示后的 messages
"temperature": kwargs.get("temperature", 0.1),
"max_tokens": kwargs.get("max_tokens", 1024),
"top_p": kwargs.get("top_p", 0.9),
}

# 5. 处理 response_format 格式转换（原有逻辑不变）
if response_format:
# 场景1：传入的是 Pydantic 模型（比如 ObserveInferenceSchema）
if isinstance(response_format, type) and issubclass(response_format, BaseModel):
payload["response_format"] = {
"type": "json_schema",
"json_schema": {
"name": response_format.__name__,
"strict": True,
"schema": response_format.model_json_schema(),
}
}
# 场景2：传入的是 json_schema 格式（但 strict 为 false）
elif isinstance(response_format, dict) and response_format.get("type") == "json_schema":
response_format["json_schema"]["strict"] = True
payload["response_format"] = response_format
# 场景3：传入的是 json_object（通用兼容）
elif isinstance(response_format, dict) and response_format.get("type") == "json_object":
payload["response_format"] = response_format
# 其他情况：兜底为 json_object
else:
payload["response_format"] = {"type": "json_object"}

# 6. 发送请求
start_time = start_inference_timer()
async with aiohttp.ClientSession() as session:
async with session.post(self.api_base, json=payload, headers=headers) as response:
if response.status != 200:
raise ValueError(f"通义千问 API 错误: {await response.text()}")
response_data = await response.json()

# 7. 解析 JSON 响应（原有逻辑不变）
if response_format:
content = response_data.get("choices", [{}])[0].get("message", {}).get("content", "")
if content:
try:
parsed_content = json.loads(content)
# ========== 核心适配逻辑 ==========
# 遍历 elements 数组，强制转换 arguments 为列表
if isinstance(parsed_content, dict) and "elements" in parsed_content:
elements = parsed_content["elements"]
if isinstance(elements, list):
for elem in elements:
# 1. 如果 arguments 字段不存在 → 初始化为空列表
if "arguments" not in elem:
elem["arguments"] = []
else:
# 2. 如果 arguments 是 None → 替换为空列表
if elem["arguments"] is None:
elem["arguments"] = []
# 3. 如果 arguments 是非列表类型（字符串/数字/布尔等）→ 包装为列表
elif not isinstance(elem["arguments"], list):
# 额外处理：如果值是 None，包装为包含空字符串的列表（按需可选）
elem["arguments"] = [elem["arguments"] if elem["arguments"] is not None else ""]
# ========== 适配结束 ==========
response_data["choices"][0]["message"]["content"] = parsed_content
except json.JSONDecodeError as e:
raise ValueError(f"千问模型返回非 JSON 内容：{content} | 错误：{str(e)}")

# 8. 调用指标回调（关键修改：包装响应为对象）
inference_time_ms = get_inference_time_ms(start_time)
if self.metrics_callback and function_name:
response_obj = HybridDict(response_data)
self.metrics_callback(response_obj, inference_time_ms, function_name)

return HybridDict(response_data)
26 changes: 19 additions & 7 deletions stagehand/main.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,6 +30,7 @@
from .metrics import StagehandFunctionName, StagehandMetrics
from .page import StagehandPage
from .utils import get_download_path, make_serializable
from .llm.qwenclient import QwenClient # 导入千问客户端

load_dotenv()

Expand Down Expand Up @@ -284,13 +285,24 @@ def __init__(
# Setup LLM client if LOCAL mode
self.llm = None
if not self.use_api:
self.llm = LLMClient(
stagehand_logger=self.logger,
api_key=self.model_api_key,
default_model=self.model_name,
metrics_callback=self._handle_llm_metrics,
**self.model_client_options,
)
# 检查是否为千问模型，使用自定义 QwenClient
if self.model_name in ["qwen-turbo", "qwen-plus", "qwen-max"]:
self.llm = QwenClient(
stagehand_logger=self.logger,
api_key=self.api_key,
model_name=self.model_name,
metrics_callback=self._handle_llm_metrics,
**self.model_client_options,
)
else:
# 其他模型使用默认 LLMClient
self.llm = LLMClient(
stagehand_logger=self.logger,
api_key=self.model_api_key,
default_model=self.model_name,
metrics_callback=self._handle_llm_metrics,
**self.model_client_options,
)

def _register_signal_handlers(self):
"""Register signal handlers for SIGINT and SIGTERM to ensure proper cleanup."""
Expand Down
4 changes: 4 additions & 0 deletions stagehand/schemas.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,10 @@ class AvailableModel(str, Enum):
CLAUDE_3_7_SONNET_LATEST = "claude-3-7-sonnet-latest"
COMPUTER_USE_PREVIEW = "computer-use-preview"
GEMINI_2_0_FLASH = "gemini-2.0-flash"
# 添加通义千问模型
QWEN_TURBO = "qwen-turbo"
QWEN_PLUS = "qwen-plus"
QWEN_MAX = "qwen-max"


class StagehandBaseModel(BaseModel):
Expand Down