Spaces:
Running
Running
Upload rag_engine.py
Browse files- rag_engine.py +480 -0
rag_engine.py
ADDED
|
@@ -0,0 +1,480 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
RAG引擎:实现传统RAG和GraphRAG的检索逻辑
|
| 3 |
+
"""
|
| 4 |
+
from typing import List, Dict, Tuple
|
| 5 |
+
# 优先使用轻量级版本(避免超过 Vercel 250MB 限制)
|
| 6 |
+
try:
|
| 7 |
+
from database_setup_lite import SimpleGraphDB, VectorDB
|
| 8 |
+
except ImportError:
|
| 9 |
+
from database_setup import SimpleGraphDB, VectorDB
|
| 10 |
+
import json
|
| 11 |
+
import requests
|
| 12 |
+
|
| 13 |
+
# LLM配置(从环境变量读取,确保安全)
|
| 14 |
+
import os
|
| 15 |
+
LLM_API_BASE = os.getenv("LLM_API_BASE", "https://api.ai-gaochao.cn/v1")
|
| 16 |
+
LLM_API_KEY = os.getenv("LLM_API_KEY", "")
|
| 17 |
+
LLM_MODEL = os.getenv("LLM_MODEL", "gemini-2.5-flash")
|
| 18 |
+
|
| 19 |
+
if not LLM_API_KEY:
|
| 20 |
+
raise ValueError("LLM_API_KEY 环境变量未设置!请在 .env 文件中设置 LLM_API_KEY")
|
| 21 |
+
|
| 22 |
+
class TraditionalRAG:
|
| 23 |
+
"""传统语义RAG"""
|
| 24 |
+
def __init__(self, vector_db: VectorDB, graph_db: SimpleGraphDB = None):
|
| 25 |
+
self.vector_db = vector_db
|
| 26 |
+
self.graph_db = graph_db # 用于限制搜索范围
|
| 27 |
+
|
| 28 |
+
def retrieve(self, query: str, product_name: str = None, style_name: str = None, n_results: int = 5) -> Dict:
|
| 29 |
+
"""语义检索(传统RAG:直接向量搜索,不利用图结构,返回片段句子)"""
|
| 30 |
+
# 传统RAG的特点:直接进行语义相似度搜索,不利用图结构
|
| 31 |
+
# 使用相同的文案数据库,但只返回相似的片段句子(而不是完整文案)
|
| 32 |
+
|
| 33 |
+
# 直接进行向量搜索(传统RAG的特点)
|
| 34 |
+
# 传统RAG限制结果数量,只返回最相关的2-3个结果
|
| 35 |
+
limited_results = min(3, n_results) # 最多返回3个结果
|
| 36 |
+
all_results = self.vector_db.search(query, n_results=limited_results * 2) # 多搜索一些,用于提取片段
|
| 37 |
+
|
| 38 |
+
# 从完整文案中提取与查询最相关的片段句子
|
| 39 |
+
processed_results = []
|
| 40 |
+
query_keywords = set(query.lower().split())
|
| 41 |
+
|
| 42 |
+
for result in all_results[:limited_results * 2]:
|
| 43 |
+
full_content = result.get("content", "")
|
| 44 |
+
if not full_content:
|
| 45 |
+
continue
|
| 46 |
+
|
| 47 |
+
# 将文案按句子分割(中文句号、英文句号、感叹号、问号)
|
| 48 |
+
import re
|
| 49 |
+
sentences = re.split(r'[。!?.!?]', full_content)
|
| 50 |
+
sentences = [s.strip() for s in sentences if s.strip()]
|
| 51 |
+
|
| 52 |
+
# 找到与查询最相关的句子片段
|
| 53 |
+
best_sentences = []
|
| 54 |
+
for sentence in sentences:
|
| 55 |
+
# 计算句子与查询的相关度(简单关键词匹配)
|
| 56 |
+
sentence_lower = sentence.lower()
|
| 57 |
+
keyword_matches = sum(1 for keyword in query_keywords if keyword in sentence_lower)
|
| 58 |
+
if keyword_matches > 0:
|
| 59 |
+
best_sentences.append((sentence, keyword_matches))
|
| 60 |
+
|
| 61 |
+
# 按相关度排序,取前2-3个最相关的句子
|
| 62 |
+
best_sentences.sort(key=lambda x: x[1], reverse=True)
|
| 63 |
+
selected_sentences = [s[0] for s in best_sentences[:3]]
|
| 64 |
+
|
| 65 |
+
# 如果没有找到相关句子,取前3个句子作为片段
|
| 66 |
+
if not selected_sentences and sentences:
|
| 67 |
+
selected_sentences = sentences[:3]
|
| 68 |
+
|
| 69 |
+
# 组合成片段(最多150字,确保有足够内容)
|
| 70 |
+
snippet = "。".join(selected_sentences)
|
| 71 |
+
if not snippet and sentences:
|
| 72 |
+
# 如果还是空的,至少取前3个句子
|
| 73 |
+
snippet = "。".join(sentences[:3])
|
| 74 |
+
if len(snippet) > 150:
|
| 75 |
+
snippet = snippet[:150] + "..."
|
| 76 |
+
elif len(snippet) < 30 and len(sentences) > 0:
|
| 77 |
+
# 如果片段太短,至少取前2-3个句子
|
| 78 |
+
snippet = "。".join(sentences[:min(3, len(sentences))])
|
| 79 |
+
if len(snippet) > 150:
|
| 80 |
+
snippet = snippet[:150] + "..."
|
| 81 |
+
|
| 82 |
+
if snippet:
|
| 83 |
+
processed_results.append({
|
| 84 |
+
"content": snippet, # 返回片段而不是完整文案
|
| 85 |
+
"full_content": full_content, # 保留完整内容用于显示
|
| 86 |
+
"metadata": result.get("metadata", {}),
|
| 87 |
+
"distance": result.get("distance", 0),
|
| 88 |
+
"is_snippet": True # 标记这是片段
|
| 89 |
+
})
|
| 90 |
+
|
| 91 |
+
if len(processed_results) >= limited_results:
|
| 92 |
+
break
|
| 93 |
+
|
| 94 |
+
# 如果结果太少,至少返回1-2个语义相似的结果
|
| 95 |
+
if len(processed_results) < 1:
|
| 96 |
+
# 如果提取片段失败,至少返回一些结果
|
| 97 |
+
for result in all_results[:max(1, limited_results)]:
|
| 98 |
+
content = result.get("content", "")
|
| 99 |
+
if content:
|
| 100 |
+
# 简单截取前150字作为片段
|
| 101 |
+
snippet = content[:150] + "..." if len(content) > 150 else content
|
| 102 |
+
processed_results.append({
|
| 103 |
+
"content": snippet,
|
| 104 |
+
"full_content": content,
|
| 105 |
+
"metadata": result.get("metadata", {}),
|
| 106 |
+
"distance": result.get("distance", 0),
|
| 107 |
+
"is_snippet": True
|
| 108 |
+
})
|
| 109 |
+
if len(processed_results) >= limited_results:
|
| 110 |
+
break
|
| 111 |
+
|
| 112 |
+
return {
|
| 113 |
+
"method": "语义检索",
|
| 114 |
+
"query": query,
|
| 115 |
+
"product": product_name,
|
| 116 |
+
"style": style_name,
|
| 117 |
+
"results": processed_results[:limited_results],
|
| 118 |
+
"retrieval_path": [
|
| 119 |
+
"向量相似度搜索(传统RAG:不利用图结构)",
|
| 120 |
+
f"找到 {len(processed_results)} 个语义相似的片段",
|
| 121 |
+
"⚠️ 局限性:只返回片段句子,没有图结构,无法找到跨品类的风格相关文案"
|
| 122 |
+
],
|
| 123 |
+
"explanation": "传统RAG直接通过语义相似度搜索相关文案,使用相同的文案数据库,但只返回与查询最相关的片段句子(而不是完整文案)。没有图结构,无法找到跨品类的风格相关文案。"
|
| 124 |
+
}
|
| 125 |
+
|
| 126 |
+
class GraphRAG:
|
| 127 |
+
"""图增强RAG"""
|
| 128 |
+
def __init__(self, graph_db: SimpleGraphDB, vector_db: VectorDB):
|
| 129 |
+
self.graph_db = graph_db
|
| 130 |
+
self.vector_db = vector_db
|
| 131 |
+
|
| 132 |
+
def retrieve(self, query: str, product_name: str = None, style_name: str = None, n_results: int = 5) -> Dict:
|
| 133 |
+
"""图增强检索"""
|
| 134 |
+
retrieval_path = []
|
| 135 |
+
retrieved_docs = []
|
| 136 |
+
|
| 137 |
+
# 步骤1: 尝试找到风格节点
|
| 138 |
+
style_node = None
|
| 139 |
+
if style_name:
|
| 140 |
+
style_node = self.graph_db.find_node_by_property("Style", "name", style_name)
|
| 141 |
+
if style_node:
|
| 142 |
+
retrieval_path.append(f"定位风格节点: {style_node['properties']['name']}")
|
| 143 |
+
|
| 144 |
+
# 步骤2: 通过风格节点找到相关文案(跨品类)
|
| 145 |
+
if style_node:
|
| 146 |
+
# 反向查找:找到连接到风格的文案节点
|
| 147 |
+
for edge in self.graph_db.edges:
|
| 148 |
+
if edge["target"] == style_node["id"] and edge["relationship"] == "HAS_STYLE":
|
| 149 |
+
copy_node = self.graph_db.nodes.get(edge["source"])
|
| 150 |
+
if copy_node and copy_node["type"] == "Copywriting":
|
| 151 |
+
content = copy_node["properties"]["content"]
|
| 152 |
+
# 获取该文案关联的产品(HAS_COPY关系:Product -> Copywriting)
|
| 153 |
+
product_id = None
|
| 154 |
+
for e in self.graph_db.edges:
|
| 155 |
+
if e["target"] == edge["source"] and e["relationship"] == "HAS_COPY":
|
| 156 |
+
product_id = e["source"]
|
| 157 |
+
break
|
| 158 |
+
|
| 159 |
+
product_info = self.graph_db.nodes.get(product_id, {}).get("properties", {})
|
| 160 |
+
retrieved_docs.append({
|
| 161 |
+
"content": content,
|
| 162 |
+
"source": "图遍历",
|
| 163 |
+
"product": product_info.get("name", "未知"),
|
| 164 |
+
"style": style_name,
|
| 165 |
+
"tag": copy_node["properties"].get("tag", ""),
|
| 166 |
+
"retrieval_reason": f"通过风格节点'{style_name}'找到的跨品类文案(来自产品:{product_info.get('name', '未知')})"
|
| 167 |
+
})
|
| 168 |
+
|
| 169 |
+
if retrieved_docs:
|
| 170 |
+
retrieval_path.append(f"通过风格节点遍历找到 {len(retrieved_docs)} 个相关文案")
|
| 171 |
+
else:
|
| 172 |
+
retrieval_path.append("未找到该风格的相关文案")
|
| 173 |
+
|
| 174 |
+
# 步骤3: 如果指定了产品,查找产品特征
|
| 175 |
+
product_features = []
|
| 176 |
+
if product_name:
|
| 177 |
+
product_node = self.graph_db.find_node_by_property("Product", "name", product_name)
|
| 178 |
+
if product_node:
|
| 179 |
+
retrieval_path.append(f"定位产品节点: {product_name}")
|
| 180 |
+
features = product_node["properties"].get("features", [])
|
| 181 |
+
keywords = product_node["properties"].get("keywords", [])
|
| 182 |
+
product_features = features + keywords
|
| 183 |
+
retrieval_path.append(f"提取产品特征: {', '.join(product_features[:5])}")
|
| 184 |
+
|
| 185 |
+
# 步骤4: 如果图检索结果不足,用向量检索补充
|
| 186 |
+
if len(retrieved_docs) < n_results:
|
| 187 |
+
vector_results = self.vector_db.search(query, n_results=n_results - len(retrieved_docs))
|
| 188 |
+
for result in vector_results:
|
| 189 |
+
# 避免重复
|
| 190 |
+
if not any(doc["content"] == result["content"] for doc in retrieved_docs):
|
| 191 |
+
retrieved_docs.append({
|
| 192 |
+
"content": result["content"],
|
| 193 |
+
"source": "向量检索补充",
|
| 194 |
+
"product": result["metadata"].get("product_id", "未知"),
|
| 195 |
+
"style": result["metadata"].get("style_id", "未知"),
|
| 196 |
+
"tag": result["metadata"].get("tag", ""),
|
| 197 |
+
"retrieval_reason": "语义���似度补充检索"
|
| 198 |
+
})
|
| 199 |
+
if vector_results:
|
| 200 |
+
retrieval_path.append(f"向量检索补充 {len(vector_results)} 个结果")
|
| 201 |
+
|
| 202 |
+
return {
|
| 203 |
+
"method": "图增强检索",
|
| 204 |
+
"query": query,
|
| 205 |
+
"product": product_name,
|
| 206 |
+
"style": style_name,
|
| 207 |
+
"product_features": product_features,
|
| 208 |
+
"results": retrieved_docs[:n_results],
|
| 209 |
+
"retrieval_path": retrieval_path,
|
| 210 |
+
"explanation": "通过图结构找到跨品类的风格相关文案,即使产品不同,但风格相通,可以借鉴文案模板。"
|
| 211 |
+
}
|
| 212 |
+
|
| 213 |
+
class RAGEngine:
|
| 214 |
+
"""RAG引擎主类"""
|
| 215 |
+
def __init__(self, graph_db: SimpleGraphDB, vector_db: VectorDB):
|
| 216 |
+
self.graph_db = graph_db
|
| 217 |
+
self.traditional_rag = TraditionalRAG(vector_db, graph_db)
|
| 218 |
+
self.graph_rag = GraphRAG(graph_db, vector_db)
|
| 219 |
+
|
| 220 |
+
def compare_retrieval(self, query: str, product_name: str = None, style_name: str = None) -> Dict:
|
| 221 |
+
"""对比传统RAG和GraphRAG的检索结果"""
|
| 222 |
+
traditional_result = self.traditional_rag.retrieve(query, product_name, style_name)
|
| 223 |
+
graph_result = self.graph_rag.retrieve(query, product_name, style_name)
|
| 224 |
+
|
| 225 |
+
return {
|
| 226 |
+
"traditional_rag": traditional_result,
|
| 227 |
+
"graph_rag": graph_result,
|
| 228 |
+
"comparison": {
|
| 229 |
+
"traditional_count": len(traditional_result["results"]),
|
| 230 |
+
"graph_count": len(graph_result["results"]),
|
| 231 |
+
"graph_cross_category": len([r for r in graph_result["results"] if r.get("source") == "图遍历"])
|
| 232 |
+
}
|
| 233 |
+
}
|
| 234 |
+
|
| 235 |
+
def generate_copywriting(self, query: str, product_name: str, style_name: str, use_graph: bool = True) -> Dict:
|
| 236 |
+
"""生成文案(使用LLM)"""
|
| 237 |
+
if use_graph:
|
| 238 |
+
retrieval_result = self.graph_rag.retrieve(query, product_name, style_name)
|
| 239 |
+
else:
|
| 240 |
+
retrieval_result = self.traditional_rag.retrieve(query, product_name, style_name)
|
| 241 |
+
|
| 242 |
+
# 获取检索到的参考文案
|
| 243 |
+
retrieved_texts = [r["content"] for r in retrieval_result["results"][:5]] # 取前5个作为参考
|
| 244 |
+
|
| 245 |
+
# 统计信息
|
| 246 |
+
cross_category_count = len([r for r in retrieval_result["results"] if r.get("source") == "图遍历"]) if use_graph else 0
|
| 247 |
+
|
| 248 |
+
# 获取产品特征(用于GraphRAG)
|
| 249 |
+
product_features = []
|
| 250 |
+
if use_graph and retrieval_result.get("product_features"):
|
| 251 |
+
product_features = retrieval_result["product_features"]
|
| 252 |
+
|
| 253 |
+
# 调用LLM生成文案
|
| 254 |
+
try:
|
| 255 |
+
llm_generated = self._call_llm_generate(
|
| 256 |
+
product_name=product_name,
|
| 257 |
+
style_name=style_name,
|
| 258 |
+
reference_texts=retrieved_texts,
|
| 259 |
+
product_features=product_features,
|
| 260 |
+
use_graph=use_graph,
|
| 261 |
+
cross_category_count=cross_category_count
|
| 262 |
+
)
|
| 263 |
+
except Exception as e:
|
| 264 |
+
print(f"LLM生成失败: {e}")
|
| 265 |
+
# 如果LLM失败,使用模板生成
|
| 266 |
+
llm_generated = self._generate_template(retrieved_texts, product_name, style_name)
|
| 267 |
+
|
| 268 |
+
# 组装最终输出
|
| 269 |
+
if use_graph and product_features:
|
| 270 |
+
features = ", ".join(product_features[:3])
|
| 271 |
+
reference_sources = ', '.join([r.get('product', '未知') for r in retrieval_result["results"][:3]])
|
| 272 |
+
generated_text = f"""基于图增强检索生成的文案:
|
| 273 |
+
|
| 274 |
+
✨ 检索策略:通过图结构找到跨品类的风格相关文案
|
| 275 |
+
📊 检索结果:找到 {len(retrieved_texts)} 个相关文案,其中 {cross_category_count} 个来自跨品类(通过风格节点关联)
|
| 276 |
+
🎯 产品特征:{features}
|
| 277 |
+
📝 参考文案来源:{reference_sources}
|
| 278 |
+
|
| 279 |
+
【{style_name}风格】{product_name}文案:
|
| 280 |
+
|
| 281 |
+
{llm_generated}
|
| 282 |
+
|
| 283 |
+
💡 说明:GraphRAG 通过风格节点找到了跨品类的参考文案(如香薰蜡烛的清冷避世风文案),即使产品不同,但风格相通,可以借鉴文案模板。"""
|
| 284 |
+
else:
|
| 285 |
+
generated_text = f"""基于传统语义检索生成的文案:
|
| 286 |
+
|
| 287 |
+
🔍 检索策略:直接通过语义相似度搜索
|
| 288 |
+
📊 检索结果:找到 {len(retrieved_texts)} 个语义相似的文案
|
| 289 |
+
⚠️ 局限性:如果数据库中没有相似内容,可能返回不相关的结果
|
| 290 |
+
|
| 291 |
+
【{style_name}风格】{product_name}文案:
|
| 292 |
+
|
| 293 |
+
{llm_generated}
|
| 294 |
+
|
| 295 |
+
💡 说明:传统 RAG 只能找到语义相似的文案,如果数据库中没有该产品的该风格文案,可能无法生成合适的文案。"""
|
| 296 |
+
|
| 297 |
+
return {
|
| 298 |
+
"generated_text": generated_text,
|
| 299 |
+
"retrieval_result": retrieval_result,
|
| 300 |
+
"method": "GraphRAG" if use_graph else "Traditional RAG"
|
| 301 |
+
}
|
| 302 |
+
|
| 303 |
+
def _call_llm_generate(self, product_name: str, style_name: str, reference_texts: List[str],
|
| 304 |
+
product_features: List[str] = None, use_graph: bool = True,
|
| 305 |
+
cross_category_count: int = 0) -> str:
|
| 306 |
+
"""调用LLM生成文案"""
|
| 307 |
+
headers = {
|
| 308 |
+
"Content-Type": "application/json",
|
| 309 |
+
"Authorization": f"Bearer {LLM_API_KEY}"
|
| 310 |
+
}
|
| 311 |
+
url = f"{LLM_API_BASE}/chat/completions"
|
| 312 |
+
|
| 313 |
+
# 构建参考文案说明
|
| 314 |
+
reference_context = ""
|
| 315 |
+
if reference_texts:
|
| 316 |
+
reference_context = "\n\n参考文案(用于学习风格和句式):\n"
|
| 317 |
+
for i, text in enumerate(reference_texts[:3], 1):
|
| 318 |
+
reference_context += f"{i}. {text}\n"
|
| 319 |
+
else:
|
| 320 |
+
reference_context = "\n\n⚠️ 注意:没有找到相关参考文案,请根据产品特征和风格要求创作。"
|
| 321 |
+
|
| 322 |
+
# 构建产品特征说明
|
| 323 |
+
features_context = ""
|
| 324 |
+
if product_features:
|
| 325 |
+
features_context = f"\n产品特征:{', '.join(product_features[:5])}"
|
| 326 |
+
|
| 327 |
+
# 构建prompt
|
| 328 |
+
if use_graph and cross_category_count > 0:
|
| 329 |
+
prompt = f"""你是一名擅长小红书文案写作的创意编辑。请根据以下信息,生成一篇适合在小红书发布的文案(200-300字,要求内容丰富、有细节感)。
|
| 330 |
+
|
| 331 |
+
产品名称:{product_name}
|
| 332 |
+
目标风格:{style_name}
|
| 333 |
+
{features_context}
|
| 334 |
+
|
| 335 |
+
{reference_context}
|
| 336 |
+
|
| 337 |
+
重要提示:
|
| 338 |
+
1. 这些参考文案来自其他产品(跨品类),但风格相同,请学习它们的句式、语气和情感表达方式
|
| 339 |
+
2. 将参考文案的风格和句式应用到目标产品上
|
| 340 |
+
3. 文案要有细节感、人情味,符合小红书用户的阅读习惯
|
| 341 |
+
4. 保持{style_name}的风格特征
|
| 342 |
+
5. 文案长度要求200-300字,要有丰富的内容和细节描述,可以包含使用场景、情感体验、产品特色等多个方面
|
| 343 |
+
6. 请确保文案完整,不要被截断,以完整的句子结尾
|
| 344 |
+
|
| 345 |
+
**必须遵守的输出格式要求:**
|
| 346 |
+
- 你必须使用中英对照格式输出文案,按段落进行中英对照
|
| 347 |
+
- 格式:中文段落(换行)English paragraph(再换行)
|
| 348 |
+
- 每个中文段落后面必须换行,然后添加对应的英文段落翻译,英文段落后再换行
|
| 349 |
+
- 示例格式:
|
| 350 |
+
这款真丝眼罩真的太舒服了,遮光效果特别好,戴上之后整个世界都安静了。
|
| 351 |
+
This silk eye mask is really comfortable, with excellent light-blocking effect. After putting it on, the whole world becomes quiet.
|
| 352 |
+
|
| 353 |
+
每天晚上睡前戴上它,就像给自己创造了一个专属的避风港。
|
| 354 |
+
Every night before sleep, putting it on is like creating a personal sanctuary for yourself.
|
| 355 |
+
|
| 356 |
+
材质柔软亲肤,完全不会压迫眼睛,真的爱了。
|
| 357 |
+
The material is soft and skin-friendly, completely non-pressuring on the eyes, I really love it.
|
| 358 |
+
- 不要只输出中文,必须每个段落都包含对应的英文翻译
|
| 359 |
+
- 可以一个段落包含多句话,然后整体翻译成英文
|
| 360 |
+
- 每个中文段落和英文段落之间必须换行,段落之间用空行分隔
|
| 361 |
+
|
| 362 |
+
请直接输出文案内容,不要包含"好的"、"没问题"等前缀,也不要使用markdown格式。只输出文案正文,确保内容完整,并且严格按照以下格式输出:中文段落(换行)English paragraph(换行)。"""
|
| 363 |
+
else:
|
| 364 |
+
prompt = f"""你是一名擅长小红书文案写作的创意编辑。请根据以下信息,生成一篇适合在小红书发布的文案(200-300字,要求内容丰富、有细节感)。
|
| 365 |
+
|
| 366 |
+
产品名称:{product_name}
|
| 367 |
+
目标风格:{style_name}
|
| 368 |
+
{features_context}
|
| 369 |
+
|
| 370 |
+
{reference_context}
|
| 371 |
+
|
| 372 |
+
重要提示:
|
| 373 |
+
1. 参考文案可能有限或不够相关,请根据产品特征和风格要求创作
|
| 374 |
+
2. 文案要有细节感、人情味,符合小红书用户的阅读习惯
|
| 375 |
+
3. 保持{style_name}的风格特征
|
| 376 |
+
4. 文案长度要求200-300字,要有丰富的内容和细节描述,可以包含使用场景、情感体验、产品特色等多个方面
|
| 377 |
+
5. 请确保文案完整,不要被截断,以完整的句子结尾
|
| 378 |
+
|
| 379 |
+
**必须遵守的输出格式要求:**
|
| 380 |
+
- 你必须使用中英对照格式输出文案,按段落进行中英对照
|
| 381 |
+
- 格式:中文段落(换行)English paragraph(再换行)
|
| 382 |
+
- 每个中文段落后面必须换行,然后添加对应的英文段落翻译,英文段落后再换行
|
| 383 |
+
- 示例格式:
|
| 384 |
+
这款真丝眼罩真的太舒服了,遮光效果特别好,戴上之后整个世界都安静了。
|
| 385 |
+
This silk eye mask is really comfortable, with excellent light-blocking effect. After putting it on, the whole world becomes quiet.
|
| 386 |
+
|
| 387 |
+
每天晚上睡前戴上它,就像给自己创造了一个专属的避风港。
|
| 388 |
+
Every night before sleep, putting it on is like creating a personal sanctuary for yourself.
|
| 389 |
+
|
| 390 |
+
材质柔软亲肤,完全不会压迫眼睛,真的爱了。
|
| 391 |
+
The material is soft and skin-friendly, completely non-pressuring on the eyes, I really love it.
|
| 392 |
+
- 不要只输出中文,必须每个段落都包含对应的英文翻译
|
| 393 |
+
- 可以一个段落包含多句话,然后整体翻译成英文
|
| 394 |
+
- 每个中��段落和英文段落之间必须换行,段落之间用空行分隔
|
| 395 |
+
|
| 396 |
+
请直接输出文案内容,不要包含"好的"、"没问题"等前缀,也不要使用markdown格式。只输出文案正文,确保内容完整,并且严格按照以下格式输出:中文段落(换行)English paragraph(换行)。"""
|
| 397 |
+
|
| 398 |
+
body = {
|
| 399 |
+
"model": LLM_MODEL,
|
| 400 |
+
"messages": [
|
| 401 |
+
{
|
| 402 |
+
"role": "system",
|
| 403 |
+
"content": "你是一名擅长文案写作的创意编辑,擅长创作小红书风格的文案。你必须使用中英对照格式输出所有文案内容,按段落进行中英对照,每个中文段落后面换行添加对应的英文翻译。格式:中文段落(换行)English paragraph"
|
| 404 |
+
},
|
| 405 |
+
{
|
| 406 |
+
"role": "user",
|
| 407 |
+
"content": prompt
|
| 408 |
+
}
|
| 409 |
+
],
|
| 410 |
+
"max_tokens": 4000, # 增加token限制以支持更长的文案(200-300字约需要800-1200 tokens,设置4000确保完整输出)
|
| 411 |
+
"temperature": 0.9
|
| 412 |
+
}
|
| 413 |
+
|
| 414 |
+
resp = requests.post(url, headers=headers, json=body, timeout=60)
|
| 415 |
+
resp.raise_for_status()
|
| 416 |
+
data = resp.json()
|
| 417 |
+
generated = data["choices"][0]["message"]["content"].strip()
|
| 418 |
+
|
| 419 |
+
# 清理生成的内容
|
| 420 |
+
# 移除常见的前缀(只移除开头的前缀,不要截断内容)
|
| 421 |
+
prefixes_to_remove = [
|
| 422 |
+
"好的,没问题!",
|
| 423 |
+
"好的,",
|
| 424 |
+
"没问题!",
|
| 425 |
+
"好的!",
|
| 426 |
+
]
|
| 427 |
+
for prefix in prefixes_to_remove:
|
| 428 |
+
if generated.startswith(prefix):
|
| 429 |
+
generated = generated[len(prefix):].strip()
|
| 430 |
+
|
| 431 |
+
# 移除markdown格式符号(但保留内容)
|
| 432 |
+
generated = generated.replace("**", "").replace("*", "").strip()
|
| 433 |
+
|
| 434 |
+
return generated
|
| 435 |
+
|
| 436 |
+
def _generate_template(self, reference_texts: List[str], product_name: str, style_name: str) -> str:
|
| 437 |
+
"""生成文案模板(简化版,实际应调用LLM)"""
|
| 438 |
+
# 如果有参考文案,提取关键句式
|
| 439 |
+
key_phrases = []
|
| 440 |
+
if reference_texts:
|
| 441 |
+
for text in reference_texts[:2]: # 只取前2个参考
|
| 442 |
+
# 提取关键句式(简单提取)
|
| 443 |
+
if "避难所" in text:
|
| 444 |
+
key_phrases.append("避难所")
|
| 445 |
+
if "安静" in text:
|
| 446 |
+
key_phrases.append("安静")
|
| 447 |
+
if "唯一" in text:
|
| 448 |
+
key_phrases.append("唯一")
|
| 449 |
+
if "绝绝子" in text:
|
| 450 |
+
key_phrases.append("绝绝子")
|
| 451 |
+
|
| 452 |
+
# 根据风格和产品生成
|
| 453 |
+
if "清冷避世风" in style_name or "深夜emo" in style_name.lower():
|
| 454 |
+
if "眼罩" in product_name:
|
| 455 |
+
if key_phrases:
|
| 456 |
+
# GraphRAG:使用参考文案的句式
|
| 457 |
+
return f"戴上眼罩的这片刻漆黑,是我在繁杂城市里唯一的{'避难所' if '避难所' in key_phrases else '避风港'}。物理意义上的关灯,也是心理上的断联。世界终于{'安静了' if '安静' in key_phrases else '静下来了'},今晚只属于我自己。"
|
| 458 |
+
else:
|
| 459 |
+
# 传统RAG:没有参考,使用通用模板
|
| 460 |
+
return f"这个{product_name}真的很不错,遮光效果好,推荐给大家使用。"
|
| 461 |
+
elif "CCD" in product_name or "相机" in product_name:
|
| 462 |
+
return "深夜拿起它,在颗粒感的画面里,所有的情绪都有了出口。低像素不是缺陷,是另一种真实。"
|
| 463 |
+
else:
|
| 464 |
+
if key_phrases:
|
| 465 |
+
return f"每一个与{product_name}的瞬间,都是我与世界的{'唯一连接' if '唯一' in key_phrases else '连接'}。"
|
| 466 |
+
else:
|
| 467 |
+
return f"这个{product_name}真的很不错,推荐给大家。"
|
| 468 |
+
elif "疯狂种草" in style_name:
|
| 469 |
+
if key_phrases and "绝绝子" in key_phrases:
|
| 470 |
+
# GraphRAG:使用参考文案的语气
|
| 471 |
+
return f"家人们谁懂啊!这个{product_name}真的绝绝子,一秒沦陷!必须人手一个!"
|
| 472 |
+
else:
|
| 473 |
+
# 传统RAG:没有参考,使用通用语气
|
| 474 |
+
return f"这个{product_name}真的很不错,推荐给大家购买!"
|
| 475 |
+
else:
|
| 476 |
+
if key_phrases:
|
| 477 |
+
return f"这个{product_name}真的很不错,{'强烈推荐' if '绝绝子' in key_phrases else '推荐'}给大家!"
|
| 478 |
+
else:
|
| 479 |
+
return f"这个{product_name}真的很不错,推荐给大家!"
|
| 480 |
+
|