戒酒的李白
263 lines
11 KiB
Python
263 lines
11 KiB
Python
import openai
|
|
import anthropic
|
|
import json
|
|
from typing import List, Dict
|
|
import os
|
|
from datetime import datetime
|
|
from utils.logger import app_logger as logging
|
|
|
|
class AIAnalyzer:
|
|
def __init__(self):
|
|
# 从环境变量获取API密钥
|
|
self.openai_key = os.getenv('OPENAI_API_KEY')
|
|
self.claude_key = os.getenv('ANTHROPIC_API_KEY')
|
|
|
|
if not self.openai_key and not self.claude_key:
|
|
raise ValueError("请至少设置一个API密钥 (OPENAI_API_KEY 或 ANTHROPIC_API_KEY)")
|
|
|
|
if self.openai_key:
|
|
openai.api_key = self.openai_key
|
|
if self.claude_key:
|
|
self.claude_client = anthropic.Anthropic(api_key=self.claude_key)
|
|
|
|
# 支持的模型列表
|
|
self.supported_models = {
|
|
# OpenAI 模型
|
|
'gpt-3.5-turbo': {'provider': 'openai', 'max_tokens': 2000, 'cost_per_1k': 0.0015},
|
|
'gpt-3.5-turbo-16k': {'provider': 'openai', 'max_tokens': 16000, 'cost_per_1k': 0.003},
|
|
'gpt-4': {'provider': 'openai', 'max_tokens': 8000, 'cost_per_1k': 0.03},
|
|
'gpt-4-32k': {'provider': 'openai', 'max_tokens': 32000, 'cost_per_1k': 0.06},
|
|
'gpt-4-turbo-preview': {'provider': 'openai', 'max_tokens': 128000, 'cost_per_1k': 0.01},
|
|
|
|
# Claude 模型
|
|
'claude-3-opus-20240229': {'provider': 'anthropic', 'max_tokens': 4000, 'cost_per_1k': 0.015},
|
|
'claude-3-sonnet-20240229': {'provider': 'anthropic', 'max_tokens': 3000, 'cost_per_1k': 0.003},
|
|
'claude-3-haiku-20240307': {'provider': 'anthropic', 'max_tokens': 2000, 'cost_per_1k': 0.0025},
|
|
'claude-2.1': {'provider': 'anthropic', 'max_tokens': 100000, 'cost_per_1k': 0.008},
|
|
'claude-2.0': {'provider': 'anthropic', 'max_tokens': 100000, 'cost_per_1k': 0.008},
|
|
'claude-instant-1.2': {'provider': 'anthropic', 'max_tokens': 100000, 'cost_per_1k': 0.0015}
|
|
}
|
|
|
|
# 不同深度的分析提示词
|
|
self.prompt_templates = {
|
|
'basic': """你是一个专业的舆情分析助手。请对每条消息进行基础的情感分析。
|
|
请按以下JSON格式返回：
|
|
{
|
|
"analysis_results": [
|
|
{
|
|
"message_id": "消息ID",
|
|
"sentiment": "情感倾向 (积极/消极/中性)",
|
|
"sentiment_score": "情感分数 (0-1)",
|
|
"keywords": ["关键词1", "关键词2"],
|
|
"key_points": "简要概述",
|
|
"influence_analysis": "基础影响分析",
|
|
"risk_level": "风险等级 (低/中/高)",
|
|
"timestamp": "分析时间戳"
|
|
}
|
|
]
|
|
}""",
|
|
'standard': """你是一个专业的舆情分析助手。请对每条消息进行标准深度的分析。
|
|
请按以下JSON格式返回：
|
|
{
|
|
"analysis_results": [
|
|
{
|
|
"message_id": "消息ID",
|
|
"sentiment": "情感倾向 (积极/消极/中性)",
|
|
"sentiment_score": "情感分数 (0-1)",
|
|
"keywords": ["关键词1", "关键词2", "关键词3"],
|
|
"key_points": "核心观点概述",
|
|
"influence_analysis": "潜在影响分析",
|
|
"risk_level": "风险等级 (低/中/高)",
|
|
"timestamp": "分析时间戳"
|
|
}
|
|
]
|
|
}""",
|
|
'deep': """你是一个专业的舆情分析助手。请对每条消息进行深度分析。
|
|
请按以下JSON格式返回：
|
|
{
|
|
"analysis_results": [
|
|
{
|
|
"message_id": "消息ID",
|
|
"sentiment": "情感倾向 (积极/消极/中性)",
|
|
"sentiment_score": "情感分数 (0-1)",
|
|
"keywords": ["关键词1", "关键词2", "关键词3", "关键词4", "关键词5"],
|
|
"key_points": "详细的核心观点分析",
|
|
"influence_analysis": "深度影响分析，包括短期和长期影响",
|
|
"risk_factors": ["风险因素1", "风险因素2", "风险因素3"],
|
|
"risk_level": "风险等级 (低/中/高)",
|
|
"suggestions": ["建议1", "建议2", "建议3"],
|
|
"timestamp": "分析时间戳"
|
|
}
|
|
]
|
|
}"""
|
|
}
|
|
|
|
async def analyze_messages(self, messages: List[Dict], batch_size: int = 50,
|
|
model_type: str = "gpt-3.5-turbo",
|
|
analysis_depth: str = "standard") -> List[Dict]:
|
|
"""分析一批消息并返回分析结果"""
|
|
try:
|
|
if model_type not in self.supported_models:
|
|
raise ValueError(f"不支持的模型类型: {model_type}")
|
|
|
|
model_info = self.supported_models[model_type]
|
|
provider = model_info['provider']
|
|
max_tokens = model_info['max_tokens']
|
|
|
|
# 根据模型类型调整批处理大小
|
|
adjusted_batch_size = min(batch_size, self._get_optimal_batch_size(model_type))
|
|
if adjusted_batch_size != batch_size:
|
|
logging.info(f"已将批处理大小从 {batch_size} 调整为 {adjusted_batch_size}")
|
|
|
|
all_results = []
|
|
total_cost = 0
|
|
|
|
# 分批处理消息
|
|
for i in range(0, len(messages), adjusted_batch_size):
|
|
batch = messages[i:i + adjusted_batch_size]
|
|
formatted_messages = []
|
|
for msg in batch:
|
|
formatted_messages.append(f"消息ID: {msg['id']}\n内容: {msg['content']}")
|
|
|
|
messages_text = "\n---\n".join(formatted_messages)
|
|
system_prompt = self.prompt_templates.get(analysis_depth, self.prompt_templates['standard'])
|
|
|
|
if provider == 'openai':
|
|
result = await self._analyze_with_openai(
|
|
messages_text,
|
|
system_prompt,
|
|
model_type,
|
|
max_tokens
|
|
)
|
|
else: # anthropic
|
|
result = await self._analyze_with_claude(
|
|
messages_text,
|
|
system_prompt,
|
|
model_type,
|
|
max_tokens
|
|
)
|
|
|
|
if result:
|
|
all_results.extend(result)
|
|
# 计算本批次成本
|
|
batch_cost = self._calculate_cost(len(messages_text), model_type)
|
|
total_cost += batch_cost
|
|
logging.info(f"批次处理完成,成本: ${batch_cost:.4f}")
|
|
|
|
logging.info(f"分析完成,总成本: ${total_cost:.4f}")
|
|
return all_results
|
|
|
|
except Exception as e:
|
|
logging.error(f"AI分析过程出错: {e}")
|
|
return []
|
|
|
|
def _get_optimal_batch_size(self, model_type: str) -> int:
|
|
"""根据模型类型获取最优批处理大小"""
|
|
model_info = self.supported_models[model_type]
|
|
max_tokens = model_info['max_tokens']
|
|
|
|
# 估算每条消息的平均token数(假设为200)
|
|
avg_tokens_per_message = 200
|
|
|
|
# 预留20%的token用于系统提示词和响应
|
|
available_tokens = int(max_tokens * 0.8)
|
|
|
|
# 计算最优批处理大小
|
|
optimal_batch_size = max(1, min(100, available_tokens // avg_tokens_per_message))
|
|
|
|
return optimal_batch_size
|
|
|
|
def _calculate_cost(self, input_length: int, model_type: str) -> float:
|
|
"""计算API调用成本"""
|
|
model_info = self.supported_models[model_type]
|
|
cost_per_1k = model_info['cost_per_1k']
|
|
|
|
# 估算token数(假设每4个字符约等于1个token)
|
|
estimated_tokens = input_length // 4
|
|
|
|
# 计算成本(美元)
|
|
cost = (estimated_tokens / 1000) * cost_per_1k
|
|
|
|
return cost
|
|
|
|
async def _analyze_with_openai(self, messages_text: str, system_prompt: str,
|
|
model: str, max_tokens: int) -> List[Dict]:
|
|
"""使用OpenAI API进行分析"""
|
|
try:
|
|
response = await openai.ChatCompletion.acreate(
|
|
model=model,
|
|
messages=[
|
|
{"role": "system", "content": system_prompt},
|
|
{"role": "user", "content": f"请分析以下消息:\n{messages_text}"}
|
|
],
|
|
temperature=0.3,
|
|
max_tokens=max_tokens,
|
|
n=1,
|
|
response_format={"type": "json_object"} # 强制JSON响应格式
|
|
)
|
|
|
|
result = json.loads(response.choices[0].message.content)
|
|
if isinstance(result, dict) and 'analysis_results' in result:
|
|
return result['analysis_results']
|
|
else:
|
|
logging.error(f"OpenAI API返回格式不正确: {response.choices[0].message.content}")
|
|
return []
|
|
|
|
except Exception as e:
|
|
logging.error(f"OpenAI API调用失败: {e}")
|
|
return []
|
|
|
|
async def _analyze_with_claude(self, messages_text: str, system_prompt: str,
|
|
model: str, max_tokens: int) -> List[Dict]:
|
|
"""使用Claude API进行分析"""
|
|
try:
|
|
response = await self.claude_client.messages.create(
|
|
model=model,
|
|
max_tokens=max_tokens,
|
|
temperature=0.3,
|
|
system=system_prompt,
|
|
messages=[
|
|
{
|
|
"role": "user",
|
|
"content": f"请分析以下消息:\n{messages_text}"
|
|
}
|
|
]
|
|
)
|
|
|
|
result = json.loads(response.content[0].text)
|
|
if isinstance(result, dict) and 'analysis_results' in result:
|
|
return result['analysis_results']
|
|
else:
|
|
logging.error(f"Claude API返回格式不正确: {response.content[0].text}")
|
|
return []
|
|
|
|
except Exception as e:
|
|
logging.error(f"Claude API调用失败: {e}")
|
|
return []
|
|
|
|
def format_analysis_for_display(self, analysis: Dict) -> Dict:
|
|
"""将分析结果格式化为前端显示格式"""
|
|
base_result = {
|
|
'id': analysis['message_id'],
|
|
'sentiment': analysis['sentiment'],
|
|
'sentiment_score': f"{float(analysis['sentiment_score']):.2%}",
|
|
'keywords': ', '.join(analysis['keywords']),
|
|
'key_points': analysis['key_points'],
|
|
'influence': analysis['influence_analysis'],
|
|
'risk_level': analysis['risk_level'],
|
|
'analysis_time': datetime.fromtimestamp(
|
|
float(analysis['timestamp'])
|
|
).strftime('%Y-%m-%d %H:%M:%S')
|
|
}
|
|
|
|
# 如果是深度分析，添加额外信息
|
|
if 'risk_factors' in analysis:
|
|
base_result.update({
|
|
'risk_factors': analysis['risk_factors'],
|
|
'suggestions': analysis['suggestions']
|
|
})
|
|
|
|
return base_result
|
|
|
|
# 创建全局AI分析器实例
|
|
ai_analyzer = AIAnalyzer() |