443 lines
15 KiB
Python
443 lines
15 KiB
Python
"""
|
||
新闻 LLM 分析模块
|
||
使用 LLM 分析新闻内容并生成交易建议
|
||
"""
|
||
import json
|
||
from typing import Dict, Any, List, Optional
|
||
from datetime import datetime
|
||
|
||
from app.utils.logger import logger
|
||
from app.news_agent.fetcher import NewsItem
|
||
from app.config import get_settings
|
||
from openai import OpenAI
|
||
|
||
|
||
class NewsAnalyzer:
|
||
"""新闻 LLM 分析器 (DeepSeek)"""
|
||
|
||
def __init__(self):
|
||
self.settings = get_settings()
|
||
self.client = None
|
||
|
||
try:
|
||
# 使用 DeepSeek API
|
||
self.client = OpenAI(
|
||
api_key=self.settings.deepseek_api_key,
|
||
base_url="https://api.deepseek.com"
|
||
)
|
||
except Exception as e:
|
||
logger.error(f"LLM 客户端初始化失败: {e}")
|
||
|
||
# 批量分析配置
|
||
self.batch_size = 10 # 每次最多分析 10 条新闻（只传标题，可以增加数量）
|
||
self.max_retries = 2
|
||
|
||
def _build_analysis_prompt(self, news_item: NewsItem) -> str:
|
||
"""构建单条新闻的分析提示词"""
|
||
|
||
prompt = f"""你是一名专业的金融新闻分析师。请分析以下新闻标题，并以 JSON 格式输出结果。
|
||
|
||
**新闻标题**: {news_item.title}
|
||
|
||
**新闻来源**: {news_item.source}
|
||
|
||
**新闻分类**: {news_item.category}
|
||
|
||
请按以下 JSON 格式输出（不要包含其他内容）:
|
||
|
||
```json
|
||
{{
|
||
"market_impact": "high/medium/low",
|
||
"impact_type": "bullish/bearish/neutral",
|
||
"sentiment": "positive/negative/neutral",
|
||
"summary": "简洁的新闻摘要（1句话，不超过50字）",
|
||
"key_points": ["关键点1", "关键点2", "关键点3"],
|
||
"trading_advice": "简洁的交易建议（1句话，不超过30字）",
|
||
"relevant_symbols": ["相关的币种或股票代码"],
|
||
"confidence": 85
|
||
}}
|
||
```
|
||
|
||
**分析要求**:
|
||
1. market_impact: 对市场的潜在影响（high/medium/low）
|
||
- **high**: 对市场或公司有**实质性、深远影响**的事件
|
||
* 改变行业格局或公司生存状态
|
||
* 监管政策重大变化（批准、禁止、调查）
|
||
* 系统性风险事件（破产、退市、重大欺诈）
|
||
- **medium**: 对价格有**短期影响**但不会改变长期趋势的事件
|
||
* 财报业绩、管理层变动、一般并购
|
||
* 评级调整、业务合作或重组
|
||
- **low**: 常规信息，影响有限
|
||
* 分析师观点、价格波动、一般评论
|
||
|
||
**判断原则**: 问自己"这条新闻会改变市场/公司的长期格局吗？" 如果会→high，如果只是短期波动→medium，如果无关紧要→low
|
||
|
||
2. impact_type: 对价格的影响方向（bullish=利好, bearish=利空, neutral=中性）
|
||
3. sentiment: 新闻情绪（positive=正面, negative=负面, neutral=中性）
|
||
4. summary: 根据标题推断并总结新闻核心内容
|
||
5. key_points: 基于标题推断3-5个关键信息点
|
||
6. trading_advice: 给出简明的交易建议
|
||
7. relevant_symbols: 根据标题列出相关的交易代码（如 BTC, ETH, NVDA, TSLA 等）
|
||
8. confidence: 分析置信度（0-100）
|
||
|
||
请只输出 JSON，不要包含其他解释。
|
||
"""
|
||
|
||
return prompt
|
||
|
||
def _build_batch_analysis_prompt(self, news_items: List[NewsItem]) -> str:
|
||
"""构建批量分析提示词"""
|
||
|
||
news_text = ""
|
||
for i, item in enumerate(news_items, 1):
|
||
news_text += f"""
|
||
--- 新闻 {i} ---
|
||
标题: {item.title}
|
||
来源: {item.source}
|
||
分类: {item.category}
|
||
---
|
||
"""
|
||
|
||
prompt = f"""你是一名专业的金融新闻分析师。请分析以下 {len(news_items)} 条新闻标题，并以 JSON 数组格式输出结果。
|
||
|
||
{news_text}
|
||
|
||
请按以下 JSON 格式输出（不要包含其他内容）:
|
||
|
||
```json
|
||
[
|
||
{{
|
||
"title": "新闻标题",
|
||
"market_impact": "high/medium/low",
|
||
"impact_type": "bullish/bearish/neutral",
|
||
"sentiment": "positive/negative/neutral",
|
||
"summary": "简洁的新闻摘要（1句话，不超过50字）",
|
||
"key_points": ["关键点1", "关键点2"],
|
||
"trading_advice": "简洁的交易建议（1句话，不超过30字）",
|
||
"relevant_symbols": ["相关代码"],
|
||
"confidence": 85
|
||
}}
|
||
]
|
||
```
|
||
|
||
**market_impact 判断标准**:
|
||
- **high**: 对市场或公司有**实质性、深远影响**的事件（改变行业格局或公司生存状态）
|
||
- **medium**: 对价格有**短期影响**但不会改变长期趋势的事件（财报、管理层变动、一般并购等）
|
||
- **low**: 常规信息，影响有限
|
||
|
||
**判断原则**: 问自己"这条新闻会改变市场/公司的长期格局吗？" 如果会→high，如果只是短期波动→medium。
|
||
|
||
请只输出 JSON 数组，不要包含其他解释。
|
||
"""
|
||
|
||
return prompt
|
||
|
||
def _parse_llm_response(self, response: str) -> Optional[Dict[str, Any]]:
|
||
"""解析 LLM 响应"""
|
||
try:
|
||
# 尝试提取 JSON
|
||
response = response.strip()
|
||
|
||
# 移除可能的 markdown 代码块标记
|
||
if response.startswith("```json"):
|
||
response = response[7:]
|
||
if response.startswith("```"):
|
||
response = response[3:]
|
||
if response.endswith("```"):
|
||
response = response[:-3]
|
||
|
||
response = response.strip()
|
||
|
||
# 解析 JSON
|
||
return json.loads(response)
|
||
|
||
except json.JSONDecodeError as e:
|
||
# 尝试修复截断的 JSON
|
||
logger.warning(f"JSON 解析失败，尝试修复: {e}")
|
||
try:
|
||
# 查找最后一个完整的对象
|
||
response = response.strip()
|
||
|
||
# 如果是数组，找到最后一个完整的对象
|
||
if response.startswith('['):
|
||
# 找到每个完整对象的结束位置
|
||
brace_count = 0
|
||
last_complete = 0
|
||
for i, char in enumerate(response):
|
||
if char == '{':
|
||
brace_count += 1
|
||
elif char == '}':
|
||
brace_count -= 1
|
||
if brace_count == 0:
|
||
last_complete = i + 1
|
||
break
|
||
|
||
if last_complete > 0:
|
||
# 提取完整的数组
|
||
fixed = response[:last_complete]
|
||
if not fixed.endswith(']'):
|
||
fixed += ']'
|
||
if not fixed.endswith('}'):
|
||
fixed += '}'
|
||
return json.loads(fixed)
|
||
except:
|
||
pass
|
||
|
||
logger.error(f"JSON 解析失败: {e}, 响应: {response[:500]}")
|
||
return None
|
||
|
||
def _parse_llm_array_response(self, response: str) -> Optional[List[Dict[str, Any]]]:
|
||
"""解析 LLM 数组响应"""
|
||
try:
|
||
# 尝试提取 JSON
|
||
response = response.strip()
|
||
|
||
# 移除可能的 markdown 代码块标记
|
||
if response.startswith("```json"):
|
||
response = response[7:]
|
||
if response.startswith("```"):
|
||
response = response[3:]
|
||
if response.endswith("```"):
|
||
response = response[:-3]
|
||
|
||
response = response.strip()
|
||
|
||
# 解析 JSON 数组
|
||
result = json.loads(response)
|
||
if isinstance(result, list):
|
||
return result
|
||
elif isinstance(result, dict) and 'title' in result:
|
||
# 如果返回单个对象，包装成数组
|
||
return [result]
|
||
return None
|
||
|
||
except json.JSONDecodeError as e:
|
||
# 尝试修复截断的 JSON 数组
|
||
logger.warning(f"JSON 数组解析失败，尝试修复: {e}")
|
||
try:
|
||
response = response.strip()
|
||
|
||
if response.startswith('['):
|
||
# 找到每个完整对象
|
||
objects = []
|
||
brace_count = 0
|
||
obj_start = -1
|
||
|
||
for i, char in enumerate(response):
|
||
if char == '{':
|
||
if obj_start == -1:
|
||
obj_start = i
|
||
brace_count += 1
|
||
elif char == '}':
|
||
brace_count -= 1
|
||
if brace_count == 0 and obj_start >= 0:
|
||
# 提取完整对象
|
||
obj_str = response[obj_start:i + 1]
|
||
try:
|
||
obj = json.loads(obj_str)
|
||
if isinstance(obj, dict) and 'title' in obj:
|
||
objects.append(obj)
|
||
except:
|
||
pass
|
||
obj_start = -1
|
||
|
||
if objects:
|
||
return objects
|
||
except:
|
||
pass
|
||
|
||
logger.error(f"JSON 数组解析失败: {e}, 响应: {response[:500]}")
|
||
return None
|
||
|
||
def analyze_single(self, news_item: NewsItem) -> Optional[Dict[str, Any]]:
|
||
"""
|
||
分析单条新闻
|
||
|
||
Args:
|
||
news_item: 新闻项
|
||
|
||
Returns:
|
||
分析结果字典或 None
|
||
"""
|
||
if not self.client:
|
||
logger.warning("LLM 客户端未初始化")
|
||
return None
|
||
|
||
try:
|
||
prompt = self._build_analysis_prompt(news_item)
|
||
|
||
for attempt in range(self.max_retries):
|
||
try:
|
||
response = self.client.chat.completions.create(
|
||
model="deepseek-chat",
|
||
messages=[
|
||
{"role": "system", "content": "你是一名专业的金融新闻分析师，擅长分析新闻标题对市场的影响。"},
|
||
{"role": "user", "content": prompt}
|
||
],
|
||
temperature=0.3,
|
||
max_tokens=1000 # 只传标题，减少输出token
|
||
)
|
||
|
||
result = self._parse_llm_response(response.choices[0].message.content)
|
||
|
||
if result:
|
||
logger.info(f"新闻分析成功: {news_item.title[:50]}... -> {result.get('market_impact')}")
|
||
return result
|
||
|
||
except Exception as e:
|
||
logger.warning(f"分析失败 (尝试 {attempt + 1}/{self.max_retries}): {e}")
|
||
|
||
logger.error(f"新闻分析失败，已达最大重试次数: {news_item.title[:50]}")
|
||
return None
|
||
|
||
except Exception as e:
|
||
logger.error(f"分析新闻时出错: {e}")
|
||
return None
|
||
|
||
def analyze_batch(self, news_items: List[NewsItem]) -> List[Optional[Dict[str, Any]]]:
|
||
"""
|
||
批量分析新闻
|
||
|
||
Args:
|
||
news_items: 新闻项列表
|
||
|
||
Returns:
|
||
分析结果列表（与输入顺序一致）
|
||
"""
|
||
if not self.client:
|
||
logger.warning("LLM 客户端未初始化")
|
||
return [None] * len(news_items)
|
||
|
||
results = []
|
||
|
||
# 分批处理
|
||
for i in range(0, len(news_items), self.batch_size):
|
||
batch = news_items[i:i + self.batch_size]
|
||
|
||
try:
|
||
prompt = self._build_batch_analysis_prompt(batch)
|
||
|
||
response = self.client.chat.completions.create(
|
||
model="deepseek-chat",
|
||
messages=[
|
||
{"role": "system", "content": "你是一名专业的金融新闻分析师，擅长分析新闻标题对市场的影响。"},
|
||
{"role": "user", "content": prompt}
|
||
],
|
||
temperature=0.3,
|
||
max_tokens=2000 # 批量分析需要更多 token
|
||
)
|
||
|
||
batch_results = self._parse_llm_array_response(response.choices[0].message.content)
|
||
|
||
if batch_results:
|
||
# 按标题匹配结果
|
||
title_to_result = {r.get('title'): r for r in batch_results if r and isinstance(r, dict)}
|
||
for item in batch:
|
||
result = title_to_result.get(item.title)
|
||
results.append(result)
|
||
if result:
|
||
logger.info(f"新闻分析成功: {item.title[:50]}... -> {result.get('market_impact')}")
|
||
else:
|
||
results.extend([None] * len(batch))
|
||
|
||
except Exception as e:
|
||
logger.error(f"批量分析失败: {e}")
|
||
results.extend([None] * len(batch))
|
||
|
||
return results
|
||
|
||
def calculate_priority(self, analysis: Dict[str, Any], quality_score: float = 0.5) -> float:
|
||
"""
|
||
根据分析结果计算优先级
|
||
|
||
Args:
|
||
analysis: LLM 分析结果
|
||
quality_score: 质量分数
|
||
|
||
Returns:
|
||
优先级分数
|
||
"""
|
||
score = 0.0
|
||
|
||
# 市场影响
|
||
impact_weights = {'high': 50, 'medium': 30, 'low': 10}
|
||
score += impact_weights.get(analysis.get('market_impact', 'low'), 10)
|
||
|
||
# 方向性（利空利好比中性重要）
|
||
if analysis.get('impact_type') in ['bullish', 'bearish']:
|
||
score += 15
|
||
|
||
# 置信度
|
||
score += (analysis.get('confidence', 50) / 100) * 10
|
||
|
||
# 质量分数
|
||
score += quality_score * 20
|
||
|
||
# 是否有相关代码
|
||
if analysis.get('relevant_symbols'):
|
||
score += 5
|
||
|
||
return score
|
||
|
||
|
||
class NewsAnalyzerSimple:
|
||
"""简化版新闻分析器（仅关键词规则，不使用 LLM）"""
|
||
|
||
def __init__(self):
|
||
pass
|
||
|
||
def analyze_single(self, news_item: NewsItem) -> Dict[str, Any]:
|
||
"""
|
||
基于规则分析新闻
|
||
|
||
Args:
|
||
news_item: 新闻项
|
||
|
||
Returns:
|
||
分析结果字典
|
||
"""
|
||
# 使用已有的影响评分
|
||
impact_score = getattr(news_item, 'impact_score', 0.0)
|
||
|
||
# 根据 impact_score 确定市场影响
|
||
if impact_score >= 1.0:
|
||
market_impact = 'high'
|
||
elif impact_score >= 0.7:
|
||
market_impact = 'medium'
|
||
else:
|
||
market_impact = 'low'
|
||
|
||
# 检查关键词确定方向
|
||
text = f"{news_item.title} {news_item.content}".lower()
|
||
|
||
bullish_keywords = ['上涨', '增长', '突破', '新高', 'bullish', 'surge', 'rally', 'gain', '批准', '合作']
|
||
bearish_keywords = ['下跌', '暴跌', '崩盘', 'ban', 'bearish', 'crash', 'plunge', 'fall', '禁令', '风险']
|
||
|
||
bullish_count = sum(1 for k in bullish_keywords if k in text)
|
||
bearish_count = sum(1 for k in bearish_keywords if k in text)
|
||
|
||
if bullish_count > bearish_count:
|
||
impact_type = 'bullish'
|
||
sentiment = 'positive'
|
||
elif bearish_count > bullish_count:
|
||
impact_type = 'bearish'
|
||
sentiment = 'negative'
|
||
else:
|
||
impact_type = 'neutral'
|
||
sentiment = 'neutral'
|
||
|
||
# 获取相关代码
|
||
relevant_symbols = list(set(getattr(news_item, 'relevant_symbols', [])))
|
||
|
||
return {
|
||
'market_impact': market_impact,
|
||
'impact_type': impact_type,
|
||
'sentiment': sentiment,
|
||
'summary': news_item.title,
|
||
'key_points': [news_item.title[:100]],
|
||
'trading_advice': getattr(news_item, 'impact_reason', '关注市场动态'),
|
||
'relevant_symbols': relevant_symbols,
|
||
'confidence': 70,
|
||
'analyzed_by': 'rules'
|
||
}
|