ccprocessor · drunkpig · Sep 25, 2025 · Sep 23, 2025 · Sep 23, 2025
diff --git a/API_zh-CN.md b/API_zh-CN.md
diff --git a/llm_web_kit/api/models/response.py b/llm_web_kit/api/models/response.py
@@ -45,6 +45,7 @@ class HTMLParseData(BaseModel):
     llm_response: Dict[str, int] = Field(default_factory=dict, description="LLM 项目打标结果")
     typical_main_html: Optional[str] = Field(None, description="解析得到的主体 HTML")
     html_target_list: List[Any] = Field(default_factory=list, description="正文候选/目标列表")
+    markdown: Optional[str] = Field(None, description="从正文HTML提取的Markdown格式内容")
 
 
 class HTMLParseResponse(BaseResponse):
@@ -68,7 +69,8 @@ class HTMLParseResponse(BaseResponse):
                         "item_id 9": 1
                     },
                     "typical_main_html": "<html></html>",
-                    "html_target_list": []
+                    "html_target_list": [],
+                    "markdown": "# Hello World"
                 },
                 "metadata": None
             }

diff --git a/llm_web_kit/api/run_server.py b/llm_web_kit/api/run_server.py
@@ -1,9 +1,10 @@
 #!/usr/bin/env python3
 """API 服务器启动脚本.
 
-用于启动 LLM Web Kit API 服务。
+用于启动 LLM Web Kit API 服务。 支持通过命令行参数 `--port` 指定端口号。
 """
 
+import argparse
 import os
 import sys
 
@@ -16,14 +17,29 @@
 
 if __name__ == "__main__":
     settings = get_settings()
+
+    # 创建参数解析器
+    parser = argparse.ArgumentParser(description="启动 LLM Web Kit API 服务器。")
+    parser.add_argument(
+        "--port",
+        type=int,
+        default=settings.port,
+        help=f"指定服务器运行的端口号 (默认: {settings.port}，从配置中读取)。"
+    )
+    args = parser.parse_args()
+
+    # 使用命令行参数指定的端口或配置中的默认端口
+    port_to_use = args.port
+
     print("启动 LLM Web Kit API 服务器...")
-    print(f"API 文档地址: http://{settings.host}:{settings.port}/docs")
-    print(f"ReDoc 文档地址: http://{settings.host}:{settings.port}/redoc")
+    print(f"服务将运行在: http://{settings.host}:{port_to_use}")
+    print(f"API 文档地址: http://{settings.host}:{port_to_use}/docs")
+    print(f"ReDoc 文档地址: http://{settings.host}:{port_to_use}/redoc")
 
     uvicorn.run(
         "llm_web_kit.api.main:app",
         host=settings.host,
-        port=settings.port,
+        port=port_to_use,
         reload=True,
         log_level=(settings.log_level or "INFO").lower()
     )
diff --git a/llm_web_kit/api/services/html_service.py b/llm_web_kit/api/services/html_service.py
@@ -5,6 +5,8 @@
 
 from typing import Any, Dict, Optional
 
+from llm_web_kit.simple import extract_content_from_main_html
+
 from ..dependencies import get_inference_service, get_logger, get_settings
 
 logger = get_logger(__name__)
@@ -67,7 +69,9 @@ async def parse_html(
             pre_data[PreDataJsonKey.LLM_RESPONSE] = llm_response
             parser = MapItemToHtmlTagsParser({})
             pre_data = parser.parse_single(pre_data)
-
+            main_html = pre_data[PreDataJsonKey.TYPICAL_MAIN_HTML]
+            mm_nlp_md = extract_content_from_main_html(url, main_html, 'mm_md', use_raw_image_url=True)
+            pre_data['markdown'] = mm_nlp_md
             # 将 PreDataJson 转为标准 dict，避免响应模型校验错误
             return dict(pre_data.items())
 

diff --git a/llm_web_kit/api/services/inference_service.py b/llm_web_kit/api/services/inference_service.py
@@ -30,8 +30,8 @@ class InferenceConfig:
     top_p: float = 0.95
     max_output_tokens: int = 8192
     tensor_parallel_size: int = 1
-    # 正式环境修改为bfloat16
-    dtype: str = 'float16'
+    # 测试环境修改为float16
+    dtype: str = 'bfloat16'
     template: bool = True
 
 
@@ -339,8 +339,8 @@ async def _init_model(self):
                 trust_remote_code=True,
                 dtype=config.dtype,
                 tensor_parallel_size=config.tensor_parallel_size,
-                # 正式环境删掉
-                max_model_len=config.max_tokens,  # 减少序列长度避免内存不足
+                # 测试环境取消注释
+                # max_model_len=config.max_tokens,  # 减少序列长度避免内存不足
             )
 
             logger.info(f'模型初始化成功: {self.model_path}')