万朱浩 / Venue-Ops

Go to a project
Toggle navigation Toggle navigation pinning
BaiFu
Committed by GitHub
34d4eeba 2 parents d2fec577 474c7657

Dev to Main: Refactor LLM Interface to Byte Stream for Improved Stability 

Dev to Main: Refactor LLM Interface to Byte Stream for Improved Stability
Inline Side-by-side
Showing 19 changed files with 315 additions and 40 deletions
@@ -5,7 +5,8 @@ Unified OpenAI-compatible LLM client for the Insight Engine, with retry support. @@ -5,7 +5,8 @@ Unified OpenAI-compatible LLM client for the Insight Engine, with retry support.
5 import os 5 import os
6 import sys 6 import sys
7 from datetime import datetime 7 from datetime import datetime
8 -from typing import Any, Dict, Optional 8 +from typing import Any, Dict, Optional, Iterator, Generator
  9 +from loguru import logger
9 10
10 from openai import OpenAI 11 from openai import OpenAI
11 12
@@ -82,6 +83,76 @@ class LLMClient: @@ -82,6 +83,76 @@ class LLMClient:
82 return self.validate_response(response.choices[0].message.content) 83 return self.validate_response(response.choices[0].message.content)
83 return "" 84 return ""
84 85
  86 + @with_retry(LLM_RETRY_CONFIG)
  87 + def stream_invoke(self, system_prompt: str, user_prompt: str, **kwargs) -> Generator[str, None, None]:
  88 + """
  89 + 流式调用LLM，逐步返回响应内容
  90 +
  91 + Args:
  92 + system_prompt: 系统提示词
  93 + user_prompt: 用户提示词
  94 + **kwargs: 额外参数（temperature, top_p等）
  95 +
  96 + Yields:
  97 + 响应文本块（str）
  98 + """
  99 + current_time = datetime.now().strftime("%Y年%m月%d日%H时%M分")
  100 + time_prefix = f"今天的实际时间是{current_time}"
  101 + if user_prompt:
  102 + user_prompt = f"{time_prefix}\n{user_prompt}"
  103 + else:
  104 + user_prompt = time_prefix
  105 + messages = [
  106 + {"role": "system", "content": system_prompt},
  107 + {"role": "user", "content": user_prompt},
  108 + ]
  109 +
  110 + allowed_keys = {"temperature", "top_p", "presence_penalty", "frequency_penalty"}
  111 + extra_params = {key: value for key, value in kwargs.items() if key in allowed_keys and value is not None}
  112 + # 强制使用流式
  113 + extra_params["stream"] = True
  114 +
  115 + timeout = kwargs.pop("timeout", self.timeout)
  116 +
  117 + try:
  118 + stream = self.client.chat.completions.create(
  119 + model=self.model_name,
  120 + messages=messages,
  121 + timeout=timeout,
  122 + **extra_params,
  123 + )
  124 +
  125 + for chunk in stream:
  126 + if chunk.choices and len(chunk.choices) > 0:
  127 + delta = chunk.choices[0].delta
  128 + if delta and delta.content:
  129 + yield delta.content
  130 + except Exception as e:
  131 + logger.error(f"流式请求失败: {str(e)}")
  132 + raise e
  133 +
  134 + def stream_invoke_to_string(self, system_prompt: str, user_prompt: str, **kwargs) -> str:
  135 + """
  136 + 流式调用LLM并安全地拼接为完整字符串（避免UTF-8多字节字符截断）
  137 +
  138 + Args:
  139 + system_prompt: 系统提示词
  140 + user_prompt: 用户提示词
  141 + **kwargs: 额外参数（temperature, top_p等）
  142 +
  143 + Returns:
  144 + 完整的响应字符串
  145 + """
  146 + # 以字节形式收集所有块
  147 + byte_chunks = []
  148 + for chunk in self.stream_invoke(system_prompt, user_prompt, **kwargs):
  149 + byte_chunks.append(chunk.encode('utf-8'))
  150 +
  151 + # 拼接所有字节，然后一次性解码
  152 + if byte_chunks:
  153 + return b''.join(byte_chunks).decode('utf-8', errors='replace')
  154 + return ""
  155 +
85 @staticmethod 156 @staticmethod
86 def validate_response(response: Optional[str]) -> str: 157 def validate_response(response: Optional[str]) -> str:
87 if response is None: 158 if response is None:
@@ -70,8 +70,8 @@ class ReportFormattingNode(BaseNode): @@ -70,8 +70,8 @@ class ReportFormattingNode(BaseNode):
70 70
71 logger.info("正在格式化最终报告") 71 logger.info("正在格式化最终报告")
72 72
73 - # 调用LLM  
74 - response = self.llm_client.invoke( 73 + # 调用LLM（流式，安全拼接UTF-8）
  74 + response = self.llm_client.stream_invoke_to_string(
75 SYSTEM_PROMPT_REPORT_FORMATTING, 75 SYSTEM_PROMPT_REPORT_FORMATTING,
76 message, 76 message,
77 ) 77 )
@@ -51,8 +51,8 @@ class ReportStructureNode(StateMutationNode): @@ -51,8 +51,8 @@ class ReportStructureNode(StateMutationNode):
51 try: 51 try:
52 logger.info(f"正在为查询生成报告结构: {self.query}") 52 logger.info(f"正在为查询生成报告结构: {self.query}")
53 53
54 - # 调用LLM  
55 - response = self.llm_client.invoke(SYSTEM_PROMPT_REPORT_STRUCTURE, self.query) 54 + # 调用LLM（流式，安全拼接UTF-8）
  55 + response = self.llm_client.stream_invoke_to_string(SYSTEM_PROMPT_REPORT_STRUCTURE, self.query)
56 56
57 # 处理响应 57 # 处理响应
58 processed_response = self.process_output(response) 58 processed_response = self.process_output(response)
@@ -65,8 +65,8 @@ class FirstSearchNode(BaseNode): @@ -65,8 +65,8 @@ class FirstSearchNode(BaseNode):
65 65
66 logger.info("正在生成首次搜索查询") 66 logger.info("正在生成首次搜索查询")
67 67
68 - # 调用LLM  
69 - response = self.llm_client.invoke(SYSTEM_PROMPT_FIRST_SEARCH, message) 68 + # 调用LLM（流式，安全拼接UTF-8）
  69 + response = self.llm_client.stream_invoke_to_string(SYSTEM_PROMPT_FIRST_SEARCH, message)
70 70
71 # 处理响应 71 # 处理响应
72 processed_response = self.process_output(response) 72 processed_response = self.process_output(response)
@@ -200,8 +200,8 @@ class ReflectionNode(BaseNode): @@ -200,8 +200,8 @@ class ReflectionNode(BaseNode):
200 200
201 logger.info("正在进行反思并生成新搜索查询") 201 logger.info("正在进行反思并生成新搜索查询")
202 202
203 - # 调用LLM  
204 - response = self.llm_client.invoke(SYSTEM_PROMPT_REFLECTION, message) 203 + # 调用LLM（流式，安全拼接UTF-8）
  204 + response = self.llm_client.stream_invoke_to_string(SYSTEM_PROMPT_REFLECTION, message)
205 205
206 # 处理响应 206 # 处理响应
207 processed_response = self.process_output(response) 207 processed_response = self.process_output(response)
@@ -99,8 +99,8 @@ class FirstSummaryNode(StateMutationNode): @@ -99,8 +99,8 @@ class FirstSummaryNode(StateMutationNode):
99 99
100 logger.info("正在生成首次段落总结") 100 logger.info("正在生成首次段落总结")
101 101
102 - # 调用LLM  
103 - response = self.llm_client.invoke(SYSTEM_PROMPT_FIRST_SUMMARY, message) 102 + # 调用LLM（流式，安全拼接UTF-8）
  103 + response = self.llm_client.stream_invoke_to_string(SYSTEM_PROMPT_FIRST_SUMMARY, message)
104 104
105 # 处理响应 105 # 处理响应
106 processed_response = self.process_output(response) 106 processed_response = self.process_output(response)
@@ -264,8 +264,8 @@ class ReflectionSummaryNode(StateMutationNode): @@ -264,8 +264,8 @@ class ReflectionSummaryNode(StateMutationNode):
264 264
265 logger.info("正在生成反思总结") 265 logger.info("正在生成反思总结")
266 266
267 - # 调用LLM  
268 - response = self.llm_client.invoke(SYSTEM_PROMPT_REFLECTION_SUMMARY, message) 267 + # 调用LLM（流式，安全拼接UTF-8）
  268 + response = self.llm_client.stream_invoke_to_string(SYSTEM_PROMPT_REFLECTION_SUMMARY, message)
269 269
270 # 处理响应 270 # 处理响应
271 processed_response = self.process_output(response) 271 processed_response = self.process_output(response)
@@ -5,7 +5,8 @@ Unified OpenAI-compatible LLM client for the Media Engine, with retry support. @@ -5,7 +5,8 @@ Unified OpenAI-compatible LLM client for the Media Engine, with retry support.
5 import os 5 import os
6 import sys 6 import sys
7 from datetime import datetime 7 from datetime import datetime
8 -from typing import Any, Dict, Optional 8 +from typing import Any, Dict, Optional, Generator
  9 +from loguru import logger
9 10
10 from openai import OpenAI 11 from openai import OpenAI
11 12
@@ -85,6 +86,76 @@ class LLMClient: @@ -85,6 +86,76 @@ class LLMClient:
85 return self.validate_response(response.choices[0].message.content) 86 return self.validate_response(response.choices[0].message.content)
86 return "" 87 return ""
87 88
  89 + @with_retry(LLM_RETRY_CONFIG)
  90 + def stream_invoke(self, system_prompt: str, user_prompt: str, **kwargs) -> Generator[str, None, None]:
  91 + """
  92 + 流式调用LLM，逐步返回响应内容
  93 +
  94 + Args:
  95 + system_prompt: 系统提示词
  96 + user_prompt: 用户提示词
  97 + **kwargs: 额外参数（temperature, top_p等）
  98 +
  99 + Yields:
  100 + 响应文本块（str）
  101 + """
  102 + current_time = datetime.now().strftime("%Y年%m月%d日%H时%M分")
  103 + time_prefix = f"今天的实际时间是{current_time}"
  104 + if user_prompt:
  105 + user_prompt = f"{time_prefix}\n{user_prompt}"
  106 + else:
  107 + user_prompt = time_prefix
  108 + messages = [
  109 + {"role": "system", "content": system_prompt},
  110 + {"role": "user", "content": user_prompt},
  111 + ]
  112 +
  113 + allowed_keys = {"temperature", "top_p", "presence_penalty", "frequency_penalty"}
  114 + extra_params = {key: value for key, value in kwargs.items() if key in allowed_keys and value is not None}
  115 + # 强制使用流式
  116 + extra_params["stream"] = True
  117 +
  118 + timeout = kwargs.pop("timeout", self.timeout)
  119 +
  120 + try:
  121 + stream = self.client.chat.completions.create(
  122 + model=self.model_name,
  123 + messages=messages,
  124 + timeout=timeout,
  125 + **extra_params,
  126 + )
  127 +
  128 + for chunk in stream:
  129 + if chunk.choices and len(chunk.choices) > 0:
  130 + delta = chunk.choices[0].delta
  131 + if delta and delta.content:
  132 + yield delta.content
  133 + except Exception as e:
  134 + logger.error(f"流式请求失败: {str(e)}")
  135 + raise e
  136 +
  137 + def stream_invoke_to_string(self, system_prompt: str, user_prompt: str, **kwargs) -> str:
  138 + """
  139 + 流式调用LLM并安全地拼接为完整字符串（避免UTF-8多字节字符截断）
  140 +
  141 + Args:
  142 + system_prompt: 系统提示词
  143 + user_prompt: 用户提示词
  144 + **kwargs: 额外参数（temperature, top_p等）
  145 +
  146 + Returns:
  147 + 完整的响应字符串
  148 + """
  149 + # 以字节形式收集所有块
  150 + byte_chunks = []
  151 + for chunk in self.stream_invoke(system_prompt, user_prompt, **kwargs):
  152 + byte_chunks.append(chunk.encode('utf-8'))
  153 +
  154 + # 拼接所有字节，然后一次性解码
  155 + if byte_chunks:
  156 + return b''.join(byte_chunks).decode('utf-8', errors='replace')
  157 + return ""
  158 +
88 @staticmethod 159 @staticmethod
89 def validate_response(response: Optional[str]) -> str: 160 def validate_response(response: Optional[str]) -> str:
90 if response is None: 161 if response is None:
@@ -68,8 +68,8 @@ class ReportFormattingNode(BaseNode): @@ -68,8 +68,8 @@ class ReportFormattingNode(BaseNode):
68 68
69 logger.info("正在格式化最终报告") 69 logger.info("正在格式化最终报告")
70 70
71 - # 调用LLM生成Markdown格式  
72 - response = self.llm_client.invoke( 71 + # 调用LLM生成Markdown格式（流式，安全拼接UTF-8）
  72 + response = self.llm_client.stream_invoke_to_string(
73 SYSTEM_PROMPT_REPORT_FORMATTING, 73 SYSTEM_PROMPT_REPORT_FORMATTING,
74 message, 74 message,
75 ) 75 )
@@ -52,7 +52,7 @@ class ReportStructureNode(StateMutationNode): @@ -52,7 +52,7 @@ class ReportStructureNode(StateMutationNode):
52 logger.info(f"正在为查询生成报告结构: {self.query}") 52 logger.info(f"正在为查询生成报告结构: {self.query}")
53 53
54 # 调用LLM 54 # 调用LLM
55 - response = self.llm_client.invoke(SYSTEM_PROMPT_REPORT_STRUCTURE, self.query) 55 + response = self.llm_client.stream_invoke_to_string(SYSTEM_PROMPT_REPORT_STRUCTURE, self.query)
56 56
57 # 处理响应 57 # 处理响应
58 processed_response = self.process_output(response) 58 processed_response = self.process_output(response)
@@ -66,7 +66,7 @@ class FirstSearchNode(BaseNode): @@ -66,7 +66,7 @@ class FirstSearchNode(BaseNode):
66 logger.info("正在生成首次搜索查询") 66 logger.info("正在生成首次搜索查询")
67 67
68 # 调用LLM 68 # 调用LLM
69 - response = self.llm_client.invoke(SYSTEM_PROMPT_FIRST_SEARCH, message) 69 + response = self.llm_client.stream_invoke_to_string(SYSTEM_PROMPT_FIRST_SEARCH, message)
70 70
71 # 处理响应 71 # 处理响应
72 processed_response = self.process_output(response) 72 processed_response = self.process_output(response)
@@ -201,7 +201,7 @@ class ReflectionNode(BaseNode): @@ -201,7 +201,7 @@ class ReflectionNode(BaseNode):
201 logger.info("正在进行反思并生成新搜索查询") 201 logger.info("正在进行反思并生成新搜索查询")
202 202
203 # 调用LLM 203 # 调用LLM
204 - response = self.llm_client.invoke(SYSTEM_PROMPT_REFLECTION, message) 204 + response = self.llm_client.stream_invoke_to_string(SYSTEM_PROMPT_REFLECTION, message)
205 205
206 # 处理响应 206 # 处理响应
207 processed_response = self.process_output(response) 207 processed_response = self.process_output(response)
@@ -99,8 +99,8 @@ class FirstSummaryNode(StateMutationNode): @@ -99,8 +99,8 @@ class FirstSummaryNode(StateMutationNode):
99 99
100 logger.info("正在生成首次段落总结") 100 logger.info("正在生成首次段落总结")
101 101
102 - # 调用LLM生成总结  
103 - response = self.llm_client.invoke( 102 + # 调用LLM生成总结（流式，安全拼接UTF-8）
  103 + response = self.llm_client.stream_invoke_to_string(
104 SYSTEM_PROMPT_FIRST_SUMMARY, 104 SYSTEM_PROMPT_FIRST_SUMMARY,
105 message, 105 message,
106 ) 106 )
@@ -267,8 +267,8 @@ class ReflectionSummaryNode(StateMutationNode): @@ -267,8 +267,8 @@ class ReflectionSummaryNode(StateMutationNode):
267 267
268 logger.info("正在生成反思总结") 268 logger.info("正在生成反思总结")
269 269
270 - # 调用LLM生成总结  
271 - response = self.llm_client.invoke( 270 + # 调用LLM生成总结（流式，安全拼接UTF-8）
  271 + response = self.llm_client.stream_invoke_to_string(
272 SYSTEM_PROMPT_REFLECTION_SUMMARY, 272 SYSTEM_PROMPT_REFLECTION_SUMMARY,
273 message, 273 message,
274 ) 274 )
@@ -5,7 +5,8 @@ Unified OpenAI-compatible LLM client for the Query Engine, with retry support. @@ -5,7 +5,8 @@ Unified OpenAI-compatible LLM client for the Query Engine, with retry support.
5 import os 5 import os
6 import sys 6 import sys
7 from datetime import datetime 7 from datetime import datetime
8 -from typing import Any, Dict, Optional 8 +from typing import Any, Dict, Optional, Generator
  9 +from loguru import logger
9 10
10 from openai import OpenAI 11 from openai import OpenAI
11 12
@@ -82,6 +83,76 @@ class LLMClient: @@ -82,6 +83,76 @@ class LLMClient:
82 return self.validate_response(response.choices[0].message.content) 83 return self.validate_response(response.choices[0].message.content)
83 return "" 84 return ""
84 85
  86 + @with_retry(LLM_RETRY_CONFIG)
  87 + def stream_invoke(self, system_prompt: str, user_prompt: str, **kwargs) -> Generator[str, None, None]:
  88 + """
  89 + 流式调用LLM，逐步返回响应内容
  90 +
  91 + Args:
  92 + system_prompt: 系统提示词
  93 + user_prompt: 用户提示词
  94 + **kwargs: 额外参数（temperature, top_p等）
  95 +
  96 + Yields:
  97 + 响应文本块（str）
  98 + """
  99 + current_time = datetime.now().strftime("%Y年%m月%d日%H时%M分")
  100 + time_prefix = f"今天的实际时间是{current_time}"
  101 + if user_prompt:
  102 + user_prompt = f"{time_prefix}\n{user_prompt}"
  103 + else:
  104 + user_prompt = time_prefix
  105 + messages = [
  106 + {"role": "system", "content": system_prompt},
  107 + {"role": "user", "content": user_prompt},
  108 + ]
  109 +
  110 + allowed_keys = {"temperature", "top_p", "presence_penalty", "frequency_penalty"}
  111 + extra_params = {key: value for key, value in kwargs.items() if key in allowed_keys and value is not None}
  112 + # 强制使用流式
  113 + extra_params["stream"] = True
  114 +
  115 + timeout = kwargs.pop("timeout", self.timeout)
  116 +
  117 + try:
  118 + stream = self.client.chat.completions.create(
  119 + model=self.model_name,
  120 + messages=messages,
  121 + timeout=timeout,
  122 + **extra_params,
  123 + )
  124 +
  125 + for chunk in stream:
  126 + if chunk.choices and len(chunk.choices) > 0:
  127 + delta = chunk.choices[0].delta
  128 + if delta and delta.content:
  129 + yield delta.content
  130 + except Exception as e:
  131 + logger.error(f"流式请求失败: {str(e)}")
  132 + raise e
  133 +
  134 + def stream_invoke_to_string(self, system_prompt: str, user_prompt: str, **kwargs) -> str:
  135 + """
  136 + 流式调用LLM并安全地拼接为完整字符串（避免UTF-8多字节字符截断）
  137 +
  138 + Args:
  139 + system_prompt: 系统提示词
  140 + user_prompt: 用户提示词
  141 + **kwargs: 额外参数（temperature, top_p等）
  142 +
  143 + Returns:
  144 + 完整的响应字符串
  145 + """
  146 + # 以字节形式收集所有块
  147 + byte_chunks = []
  148 + for chunk in self.stream_invoke(system_prompt, user_prompt, **kwargs):
  149 + byte_chunks.append(chunk.encode('utf-8'))
  150 +
  151 + # 拼接所有字节，然后一次性解码
  152 + if byte_chunks:
  153 + return b''.join(byte_chunks).decode('utf-8', errors='replace')
  154 + return ""
  155 +
85 @staticmethod 156 @staticmethod
86 def validate_response(response: Optional[str]) -> str: 157 def validate_response(response: Optional[str]) -> str:
87 if response is None: 158 if response is None:
@@ -68,8 +68,8 @@ class ReportFormattingNode(BaseNode): @@ -68,8 +68,8 @@ class ReportFormattingNode(BaseNode):
68 68
69 logger.info("正在格式化最终报告") 69 logger.info("正在格式化最终报告")
70 70
71 - # 调用LLM生成Markdown格式  
72 - response = self.llm_client.invoke( 71 + # 调用LLM生成Markdown格式（流式，安全拼接UTF-8）
  72 + response = self.llm_client.stream_invoke_to_string(
73 SYSTEM_PROMPT_REPORT_FORMATTING, 73 SYSTEM_PROMPT_REPORT_FORMATTING,
74 message, 74 message,
75 ) 75 )
@@ -52,7 +52,7 @@ class ReportStructureNode(StateMutationNode): @@ -52,7 +52,7 @@ class ReportStructureNode(StateMutationNode):
52 logger.info(f"正在为查询生成报告结构: {self.query}") 52 logger.info(f"正在为查询生成报告结构: {self.query}")
53 53
54 # 调用LLM 54 # 调用LLM
55 - response = self.llm_client.invoke(SYSTEM_PROMPT_REPORT_STRUCTURE, self.query) 55 + response = self.llm_client.stream_invoke_to_string(SYSTEM_PROMPT_REPORT_STRUCTURE, self.query)
56 56
57 # 处理响应 57 # 处理响应
58 processed_response = self.process_output(response) 58 processed_response = self.process_output(response)
@@ -66,7 +66,7 @@ class FirstSearchNode(BaseNode): @@ -66,7 +66,7 @@ class FirstSearchNode(BaseNode):
66 logger.info("正在生成首次搜索查询") 66 logger.info("正在生成首次搜索查询")
67 67
68 # 调用LLM 68 # 调用LLM
69 - response = self.llm_client.invoke(SYSTEM_PROMPT_FIRST_SEARCH, message) 69 + response = self.llm_client.stream_invoke_to_string(SYSTEM_PROMPT_FIRST_SEARCH, message)
70 70
71 # 处理响应 71 # 处理响应
72 processed_response = self.process_output(response) 72 processed_response = self.process_output(response)
@@ -201,7 +201,7 @@ class ReflectionNode(BaseNode): @@ -201,7 +201,7 @@ class ReflectionNode(BaseNode):
201 logger.info("正在进行反思并生成新搜索查询") 201 logger.info("正在进行反思并生成新搜索查询")
202 202
203 # 调用LLM 203 # 调用LLM
204 - response = self.llm_client.invoke(SYSTEM_PROMPT_REFLECTION, message) 204 + response = self.llm_client.stream_invoke_to_string(SYSTEM_PROMPT_REFLECTION, message)
205 205
206 # 处理响应 206 # 处理响应
207 processed_response = self.process_output(response) 207 processed_response = self.process_output(response)
@@ -99,8 +99,8 @@ class FirstSummaryNode(StateMutationNode): @@ -99,8 +99,8 @@ class FirstSummaryNode(StateMutationNode):
99 99
100 logger.info("正在生成首次段落总结") 100 logger.info("正在生成首次段落总结")
101 101
102 - # 调用LLM生成总结  
103 - response = self.llm_client.invoke( 102 + # 调用LLM生成总结（流式，安全拼接UTF-8）
  103 + response = self.llm_client.stream_invoke_to_string(
104 SYSTEM_PROMPT_FIRST_SUMMARY, 104 SYSTEM_PROMPT_FIRST_SUMMARY,
105 message, 105 message,
106 ) 106 )
@@ -267,8 +267,8 @@ class ReflectionSummaryNode(StateMutationNode): @@ -267,8 +267,8 @@ class ReflectionSummaryNode(StateMutationNode):
267 267
268 logger.info("正在生成反思总结") 268 logger.info("正在生成反思总结")
269 269
270 - # 调用LLM生成总结  
271 - response = self.llm_client.invoke( 270 + # 调用LLM生成总结（流式，安全拼接UTF-8）
  271 + response = self.llm_client.stream_invoke_to_string(
272 SYSTEM_PROMPT_REFLECTION_SUMMARY, 272 SYSTEM_PROMPT_REFLECTION_SUMMARY,
273 message, 273 message,
274 ) 274 )
@@ -4,7 +4,8 @@ Unified OpenAI-compatible LLM client for the Report Engine, with retry support. @@ -4,7 +4,8 @@ Unified OpenAI-compatible LLM client for the Report Engine, with retry support.
4 4
5 import os 5 import os
6 import sys 6 import sys
7 -from typing import Any, Dict, Optional 7 +from typing import Any, Dict, Optional, Generator
  8 +from loguru import logger
8 9
9 from openai import OpenAI 10 from openai import OpenAI
10 11
@@ -75,6 +76,70 @@ class LLMClient: @@ -75,6 +76,70 @@ class LLMClient:
75 return self.validate_response(response.choices[0].message.content) 76 return self.validate_response(response.choices[0].message.content)
76 return "" 77 return ""
77 78
  79 + @with_retry(LLM_RETRY_CONFIG)
  80 + def stream_invoke(self, system_prompt: str, user_prompt: str, **kwargs) -> Generator[str, None, None]:
  81 + """
  82 + 流式调用LLM，逐步返回响应内容
  83 +
  84 + Args:
  85 + system_prompt: 系统提示词
  86 + user_prompt: 用户提示词
  87 + **kwargs: 额外参数（temperature, top_p等）
  88 +
  89 + Yields:
  90 + 响应文本块（str）
  91 + """
  92 + messages = [
  93 + {"role": "system", "content": system_prompt},
  94 + {"role": "user", "content": user_prompt},
  95 + ]
  96 +
  97 + allowed_keys = {"temperature", "top_p", "presence_penalty", "frequency_penalty"}
  98 + extra_params = {key: value for key, value in kwargs.items() if key in allowed_keys and value is not None}
  99 + # 强制使用流式
  100 + extra_params["stream"] = True
  101 +
  102 + timeout = kwargs.pop("timeout", self.timeout)
  103 +
  104 + try:
  105 + stream = self.client.chat.completions.create(
  106 + model=self.model_name,
  107 + messages=messages,
  108 + timeout=timeout,
  109 + **extra_params,
  110 + )
  111 +
  112 + for chunk in stream:
  113 + if chunk.choices and len(chunk.choices) > 0:
  114 + delta = chunk.choices[0].delta
  115 + if delta and delta.content:
  116 + yield delta.content
  117 + except Exception as e:
  118 + logger.error(f"流式请求失败: {str(e)}")
  119 + raise e
  120 +
  121 + def stream_invoke_to_string(self, system_prompt: str, user_prompt: str, **kwargs) -> str:
  122 + """
  123 + 流式调用LLM并安全地拼接为完整字符串（避免UTF-8多字节字符截断）
  124 +
  125 + Args:
  126 + system_prompt: 系统提示词
  127 + user_prompt: 用户提示词
  128 + **kwargs: 额外参数（temperature, top_p等）
  129 +
  130 + Returns:
  131 + 完整的响应字符串
  132 + """
  133 + # 以字节形式收集所有块
  134 + byte_chunks = []
  135 + for chunk in self.stream_invoke(system_prompt, user_prompt, **kwargs):
  136 + byte_chunks.append(chunk.encode('utf-8'))
  137 +
  138 + # 拼接所有字节，然后一次性解码
  139 + if byte_chunks:
  140 + return b''.join(byte_chunks).decode('utf-8', errors='replace')
  141 + return ""
  142 +
78 @staticmethod 143 @staticmethod
79 def validate_response(response: Optional[str]) -> str: 144 def validate_response(response: Optional[str]) -> str:
80 if response is None: 145 if response is None:
@@ -60,7 +60,7 @@ class HTMLGenerationNode(StateMutationNode): @@ -60,7 +60,7 @@ class HTMLGenerationNode(StateMutationNode):
60 message = json.dumps(llm_input, ensure_ascii=False, indent=2) 60 message = json.dumps(llm_input, ensure_ascii=False, indent=2)
61 61
62 # 调用LLM生成HTML 62 # 调用LLM生成HTML
63 - response = self.llm_client.invoke(SYSTEM_PROMPT_HTML_GENERATION, message) 63 + response = self.llm_client.stream_invoke_to_string(SYSTEM_PROMPT_HTML_GENERATION, message)
64 64
65 # 处理响应（简化版） 65 # 处理响应（简化版）
66 processed_response = self.process_output(response) 66 processed_response = self.process_output(response)
@@ -115,7 +115,7 @@ class TemplateSelectionNode(BaseNode): @@ -115,7 +115,7 @@ class TemplateSelectionNode(BaseNode):
115 请根据查询内容、报告内容和论坛日志的具体情况，选择最合适的模板。""" 115 请根据查询内容、报告内容和论坛日志的具体情况，选择最合适的模板。"""
116 116
117 # 调用LLM 117 # 调用LLM
118 - response = self.llm_client.invoke(SYSTEM_PROMPT_TEMPLATE_SELECTION, user_message) 118 + response = self.llm_client.stream_invoke_to_string(SYSTEM_PROMPT_TEMPLATE_SELECTION, user_message)
119 119
120 # 检查响应是否为空 120 # 检查响应是否为空
121 if not response or not response.strip(): 121 if not response or not response.strip():
@@ -6,10 +6,7 @@ Forum日志读取工具 @@ -6,10 +6,7 @@ Forum日志读取工具
6 import re 6 import re
7 from pathlib import Path 7 from pathlib import Path
8 from typing import Optional, List, Dict 8 from typing import Optional, List, Dict
9 -import logging  
10 -  
11 -logger = logging.getLogger(__name__)  
12 - 9 +from loguru import logger
13 10
14 def get_latest_host_speech(log_dir: str = "logs") -> Optional[str]: 11 def get_latest_host_speech(log_dir: str = "logs") -> Optional[str]:
15 """ 12 """