刘国栋
223 lines
7.2 KiB
Python
223 lines
7.2 KiB
Python
"""
|
||
内容质量评分系统
|
||
自动评估内容是否符合 GEO 原则，提供改进建议
|
||
"""
|
||
from typing import Dict, List, Optional
|
||
from langchain_core.prompts import PromptTemplate
|
||
from langchain_core.output_parsers import StrOutputParser
|
||
import json
|
||
import re
|
||
|
||
|
||
class ContentScorer:
|
||
"""内容质量评分器"""
|
||
|
||
def __init__(self):
|
||
self.scoring_prompt_template = """
|
||
你是一名 GEO（生成式引擎优化）内容质量评估专家。请对以下内容进行全面评估，并给出详细的评分和改进建议。
|
||
|
||
【内容】
|
||
{content}
|
||
|
||
【品牌】{brand}
|
||
【优势】{advantages}
|
||
【平台】{platform}
|
||
|
||
【评估维度】
|
||
请从以下维度进行评估（每个维度 0-25 分，总分 100 分）：
|
||
|
||
1. **结构化程度**（25分）
|
||
- 是否有清晰的标题层级？
|
||
- 是否包含清单、列表、FAQ 等结构化元素？
|
||
- 内容层次是否清晰？
|
||
- 是否有结论摘要？
|
||
|
||
2. **品牌提及质量**（25分）
|
||
- 品牌提及次数是否合适（2-4次）？
|
||
- 品牌提及位置是否靠前（前1/3优先）？
|
||
- 品牌提及是否自然（先通用标准，再品牌适用）？
|
||
- 品牌与内容的关联度如何？
|
||
|
||
3. **内容权威性**（25分）
|
||
- 是否有数据支撑或案例引用？
|
||
- 是否有评估维度或选择标准？
|
||
- 是否避免编造数据（使用占位建议）？
|
||
- 内容是否专业可信？
|
||
|
||
4. **可引用性**（25分）
|
||
- 信息密度是否高？
|
||
- 结论是否先行？
|
||
- 是否容易被 AI 提取和引用？
|
||
- 是否符合目标平台的格式要求？
|
||
|
||
【输出格式】
|
||
请严格按照以下 JSON 格式输出，不要添加任何其他内容：
|
||
|
||
{{
|
||
"scores": {{
|
||
"structure": <结构化得分 0-25>,
|
||
"brand_mention": <品牌提及得分 0-25>,
|
||
"authority": <权威性得分 0-25>,
|
||
"citations": <可引用性得分 0-25>,
|
||
"total": <总分 0-100>
|
||
}},
|
||
"details": {{
|
||
"structure": "<结构化评估详情>",
|
||
"brand_mention": "<品牌提及评估详情>",
|
||
"authority": "<权威性评估详情>",
|
||
"citations": "<可引用性评估详情>"
|
||
}},
|
||
"improvements": [
|
||
"<改进建议1>",
|
||
"<改进建议2>",
|
||
"<改进建议3>"
|
||
],
|
||
"strengths": [
|
||
"<优点1>",
|
||
"<优点2>"
|
||
]
|
||
}}
|
||
|
||
【开始评估】
|
||
"""
|
||
|
||
def score_content(self, content: str, brand: str, advantages: str,
|
||
platform: str, llm_chain) -> Dict:
|
||
"""
|
||
对内容进行质量评分
|
||
|
||
Args:
|
||
content: 要评分的内容
|
||
brand: 品牌名称
|
||
advantages: 品牌优势
|
||
platform: 发布平台
|
||
llm_chain: LangChain 链对象
|
||
|
||
Returns:
|
||
包含评分、详情和改进建议的字典
|
||
"""
|
||
try:
|
||
prompt = PromptTemplate.from_template(self.scoring_prompt_template)
|
||
chain = prompt | llm_chain | StrOutputParser()
|
||
|
||
result = chain.invoke({
|
||
"content": content,
|
||
"brand": brand,
|
||
"advantages": advantages,
|
||
"platform": platform
|
||
})
|
||
|
||
# 尝试解析 JSON
|
||
score_data = self._parse_score_result(result)
|
||
|
||
return score_data
|
||
|
||
except Exception as e:
|
||
# 如果评分失败，返回默认评分
|
||
return {
|
||
"scores": {
|
||
"structure": 0,
|
||
"brand_mention": 0,
|
||
"authority": 0,
|
||
"citations": 0,
|
||
"total": 0
|
||
},
|
||
"details": {
|
||
"structure": f"评分失败：{str(e)}",
|
||
"brand_mention": "",
|
||
"authority": "",
|
||
"citations": ""
|
||
},
|
||
"improvements": ["评分系统暂时无法评估此内容，请手动检查"],
|
||
"strengths": []
|
||
}
|
||
|
||
def _parse_score_result(self, result: str) -> Dict:
|
||
"""解析评分结果"""
|
||
# 尝试提取 JSON
|
||
json_match = re.search(r'\{.*\}', result, re.DOTALL)
|
||
if json_match:
|
||
try:
|
||
score_data = json.loads(json_match.group())
|
||
# 验证数据结构
|
||
if "scores" in score_data and "total" in score_data["scores"]:
|
||
return score_data
|
||
except json.JSONDecodeError:
|
||
pass
|
||
|
||
# 如果无法解析 JSON，尝试从文本中提取信息
|
||
return self._extract_scores_from_text(result)
|
||
|
||
def _extract_scores_from_text(self, text: str) -> Dict:
|
||
"""从文本中提取评分信息（备用方案）"""
|
||
# 尝试提取总分
|
||
total_match = re.search(r'总分[：:]\s*(\d+)', text)
|
||
total_score = int(total_match.group(1)) if total_match else 0
|
||
|
||
# 简单分配分数（如果无法精确提取）
|
||
avg_score = total_score // 4 if total_score > 0 else 0
|
||
|
||
return {
|
||
"scores": {
|
||
"structure": avg_score,
|
||
"brand_mention": avg_score,
|
||
"authority": avg_score,
|
||
"citations": avg_score,
|
||
"total": total_score
|
||
},
|
||
"details": {
|
||
"structure": "无法解析详细评分",
|
||
"brand_mention": "无法解析详细评分",
|
||
"authority": "无法解析详细评分",
|
||
"citations": "无法解析详细评分"
|
||
},
|
||
"improvements": ["请检查内容是否符合 GEO 原则"],
|
||
"strengths": []
|
||
}
|
||
|
||
def get_score_level(self, total_score: int) -> tuple:
|
||
"""
|
||
根据总分返回等级和颜色
|
||
|
||
Returns:
|
||
(等级名称, 颜色代码)
|
||
"""
|
||
if total_score >= 90:
|
||
return ("优秀", "#10B981") # 绿色
|
||
elif total_score >= 75:
|
||
return ("良好", "#3B82F6") # 蓝色
|
||
elif total_score >= 60:
|
||
return ("中等", "#F59E0B") # 橙色
|
||
else:
|
||
return ("需改进", "#EF4444") # 红色
|
||
|
||
def get_quick_assessment(self, content: str, brand: str) -> Dict:
|
||
"""
|
||
快速评估（不调用 LLM，基于规则）
|
||
用于在 LLM 评分前提供初步评估
|
||
"""
|
||
assessment = {
|
||
"has_title": bool(re.search(r'^#+\s+|^标题|^##', content, re.MULTILINE)),
|
||
"has_list": bool(re.search(r'[-*•]\s+|^\d+[\.\)]\s+', content, re.MULTILINE)),
|
||
"has_faq": bool(re.search(r'FAQ|常见问题|Q[：:]|问[：:]', content, re.IGNORECASE)),
|
||
"brand_count": len(re.findall(re.escape(brand), content, re.IGNORECASE)),
|
||
"word_count": len(content)
|
||
}
|
||
|
||
# 计算初步分数
|
||
quick_score = 0
|
||
if assessment["has_title"]:
|
||
quick_score += 5
|
||
if assessment["has_list"]:
|
||
quick_score += 5
|
||
if assessment["has_faq"]:
|
||
quick_score += 5
|
||
if 2 <= assessment["brand_count"] <= 4:
|
||
quick_score += 10
|
||
elif assessment["brand_count"] > 4:
|
||
quick_score += 5
|
||
|
||
assessment["quick_score"] = min(quick_score, 30) # 最高30分（快速评估）
|
||
|
||
return assessment
|