合并分支

* PERF 修复prompt的label选择
* Merge branch 'main' into cluster-gemini
* PERF Gemini推理增加SystemInstruct
* PERF 增加上下文商品信息参数

---------

Merge request URL: https://g-ldyi2063.coding.net/p/dev/d/modalDeploy/git/merge/4867
Co-authored-by: 康宇佳

src/BowongModalFunctions/utils/HTTPUtils.py

@@ -7,7 +7,7 @@ import backoff
 import httpx
 import asyncio
 
-from google.genai.types import GenerateContentResponse
+from google.genai.types import GenerateContentResponse, ContentUnion
 from loguru import logger
 import aiofiles
 from pathlib import Path
@@ -143,6 +143,11 @@ class GoogleAuthUtils:
         def safety_settings(self) -> Optional[List[types.SafetySetting]]:
             return self.generation_config.safety_settings
 
+        @computed_field(alias="systemInstruction")
+        @property
+        def system_instruction(self) -> Optional[ContentUnion]:
+            return self.generation_config.system_instruction
+
     class GoogleGenaiClient(BaseModel):
         cloudflare_project_id: str
         cloudflare_gateway_id: str
@@ -160,7 +165,7 @@ class GoogleAuthUtils:
                              timeout: int = 30) -> tuple[dict[Any, Any], int] | tuple[GenerateContentResponse, int]:
             parameter_model = GoogleAuthUtils.VertexAIRequestModel(contents=contents, generationConfig=config)
             json_body = parameter_model.model_dump_json(indent=2, exclude_none=True, by_alias=True,
-                                                        exclude={"generation_config": {"safety_settings"}})
+                                                        exclude={"generation_config": ["safety_settings", "system_instruction"]})
             logger.info(json_body)
             url = f"{self.gateway_url}/{model_id}:generateContent"
             logger.info(f"Authorization : Bearer {self.access_token}")

src/cluster/video_apps/hls_slice_inference.py

@@ -94,11 +94,18 @@ with downloader_image.imports():
             return {
                     "temperature": c["temperature"],
                     "top_p": c["top_p"],
-                    "safety_settings": c["safety_settings"]
+                    "safety_settings": c["safety_settings"],
-                },c
+                    "system_instruction": c["system_instruction"]
+                },{
+                "temperature": c["temperature"],
+                "top_p": c["top_p"],
+                "response_mime_type": c["response_mime_type"],
+                "response_schema": c["response_schema"],
+                "safety_settings": c["safety_settings"]
+            }
 
         # 动态Prompt, langfuse获取失败使用默认值
-        IMAGE_PRODUCT_IDENTIFICATION_PROMPT = langfuse.get_prompt("Gemini自动切条/商品识别", type="text", label="latest")
+        IMAGE_PRODUCT_IDENTIFICATION_PROMPT = langfuse.get_prompt("Gemini自动切条/商品识别", type="text", label="production")
         if IMAGE_PRODUCT_IDENTIFICATION_PROMPT:
             prompt_config = IMAGE_PRODUCT_IDENTIFICATION_PROMPT.config
             first_stage_generate_config, first_stage_correct_config = split_prompt_config(prompt_config)
@@ -179,10 +186,17 @@ with downloader_image.imports():
             first_stage_generate_config = {
                 "temperature": first_stage_correct_config["temperature"],
                 "top_p": first_stage_correct_config["top_p"],
-                "safety_settings": first_stage_correct_config["safety_settings"]
+                "safety_settings": first_stage_correct_config["safety_settings"],
+                "system_instruction":{
+                    "parts": [
+                        {
+                            "text": "你是商品识别专家，任务是从商品网格图片中精准识别商品、提取特征，输出为用户定义的json格式"
+                        }
+                    ]
+                }
             }
 
-        VIDEO_TIMELINE_ANALYSIS_PROMPT = langfuse.get_prompt("Gemini自动切条/视频时间点识别", type="text", label="latest")
+        VIDEO_TIMELINE_ANALYSIS_PROMPT = langfuse.get_prompt("Gemini自动切条/视频时间点识别", type="text", label="production")
         if VIDEO_TIMELINE_ANALYSIS_PROMPT:
             prompt_config = VIDEO_TIMELINE_ANALYSIS_PROMPT.config
             second_stage_generate_config, second_stage_correct_config = split_prompt_config(prompt_config)
@@ -277,7 +291,14 @@ with downloader_image.imports():
             second_stage_generate_config = {
                 "temperature": second_stage_correct_config["temperature"],
                 "top_p": second_stage_correct_config["top_p"],
-                "safety_settings": second_stage_correct_config["safety_settings"]
+                "safety_settings": second_stage_correct_config["safety_settings"],
+                "system_instruction":{
+                    "parts": [
+                        {
+                            "text": "你是直播带货视频识别专家，任务是根据商品特征从视频中高效精准地识别出视频中每个商品出现的时间段以及内容类型，输出为用户定义的json格式"
+                        }
+                    ]
+                }
             }
 
 
@@ -308,7 +329,7 @@ with downloader_image.imports():
                                                                                       types.Part.from_text(
                                                                                           text="<prompt>"
                                                                                                "<instruction>"
-                                                                                               "请格式化以下一段非标准json格式字符串为json标准格式  \n{0}"
+                                                                                               "格式化以下一段有问题的json, json可能存在格式错误字段错误以及部分缺失，修复后输出，若product字段中存在双引号等需转义字符需要转义 \n{0}"
                                                                                                "</instruction>"
                                                                                                "</prompt>".format(
                                                                                                 json_like_str