init: RAG 知识库服务初始提交
- 后端 API（Flask + Gunicorn） - RAG 引擎（混合检索 + 云端 Reranker + 引用溯源） - 文档解析（MinerU + 多格式支持） - Docker 生产部署配置 - 排除前端项目、敏感配置、模型文件
This commit is contained in:
lacerate551
271
core/agentic_query.py
Normal file
271
core/agentic_query.py
Normal file
@@ -0,0 +1,271 @@
|
||||
"""
|
||||
Agentic RAG - 查询重写 Mixin
|
||||
|
||||
包含查询改写、实体补全、专业术语映射等方法
|
||||
"""
|
||||
|
||||
import re
|
||||
import logging
|
||||
|
||||
from .agentic_base import logger, MODEL
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class QueryRewriteMixin:
|
||||
"""查询重写方法"""
|
||||
|
||||
def _rewrite_query(self, query: str, history: list = None,
|
||||
strategy: str = "professional") -> str:
|
||||
"""
|
||||
增强版查询重写：将口语化表达转为专业术语
|
||||
|
||||
Args:
|
||||
query: 原始查询
|
||||
history: 对话历史（用于实体补全）
|
||||
strategy: 重写策略
|
||||
- professional: 口语化→专业术语
|
||||
- expand: 扩展关键词
|
||||
- clarify: 消歧义
|
||||
- entity: 实体补全
|
||||
|
||||
Returns:
|
||||
str: 重写后的查询
|
||||
"""
|
||||
# 尝试多种策略组合
|
||||
rewritten = query
|
||||
|
||||
# 策略1: 口语化→专业术语映射
|
||||
if strategy in ["professional", "all"]:
|
||||
rewritten = self._apply_professional_mapping(rewritten)
|
||||
|
||||
# 策略2: 实体补全（利用对话历史）
|
||||
if strategy in ["entity", "all"] and history:
|
||||
rewritten = self._complete_entities(rewritten, history)
|
||||
|
||||
# 策略3: LLM 深度重写（仅在需要时调用）
|
||||
if strategy in ["professional", "all"]:
|
||||
llm_rewritten = self._llm_rewrite(rewritten)
|
||||
if llm_rewritten and len(llm_rewritten) > len(rewritten) * 0.5:
|
||||
rewritten = llm_rewritten
|
||||
|
||||
return rewritten
|
||||
|
||||
def _apply_professional_mapping(self, query: str) -> str:
|
||||
"""应用口语化→专业术语映射"""
|
||||
TERM_MAPPING = {
|
||||
"报销": "差旅报销 费用报销 报销审批",
|
||||
"请假": "休假申请 请假审批 考勤管理",
|
||||
"加班": "加班申请 工时管理 加班审批",
|
||||
"工资": "薪酬管理 工资发放 薪资结构",
|
||||
"合同": "合同管理 合同签署 合同审批",
|
||||
"流程": "审批流程 业务流程 工作流",
|
||||
"制度": "管理制度 规章制度 企业规范",
|
||||
"规定": "管理规定 制度规定 政策要求",
|
||||
"几天": "时限 审批时限 办理时限",
|
||||
"多久": "处理时效 审批周期 办理周期",
|
||||
"多少": "标准 额度 限额 标准",
|
||||
"能不能": "是否允许 是否可以 权限",
|
||||
"人事": "人力资源 HR 人力部门",
|
||||
"财务": "财务部 财务部门 财务管理",
|
||||
"技术": "技术部 研发部 IT部门",
|
||||
}
|
||||
|
||||
result = query
|
||||
for colloquial, professional in TERM_MAPPING.items():
|
||||
if colloquial in query:
|
||||
result = result.replace(colloquial, f"{colloquial} {professional.split()[0]}")
|
||||
|
||||
return result
|
||||
|
||||
def _complete_entities(self, query: str, history: list) -> str:
|
||||
"""实体补全：利用对话历史补充缺失的实体"""
|
||||
if not history:
|
||||
return query
|
||||
|
||||
# 图片指代识别
|
||||
image_reference = self._detect_image_reference(query, history)
|
||||
if image_reference:
|
||||
return image_reference
|
||||
|
||||
# 获取最近用户消息
|
||||
last_user_msg = None
|
||||
for msg in reversed(history):
|
||||
if msg.get("role") == "user":
|
||||
last_user_msg = msg.get("content", "")
|
||||
break
|
||||
|
||||
if not last_user_msg:
|
||||
return query
|
||||
|
||||
# 检查当前查询是否缺少主语
|
||||
BUSINESS_KEYWORDS = ["报销", "出差", "请假", "工资", "合同", "审批", "流程",
|
||||
"制度", "规定", "标准", "金额", "时间"]
|
||||
|
||||
has_subject = any(kw in query for kw in BUSINESS_KEYWORDS)
|
||||
|
||||
if not has_subject:
|
||||
try:
|
||||
import jieba
|
||||
entities = []
|
||||
for word in jieba.cut(last_user_msg):
|
||||
word = word.strip()
|
||||
if len(word) >= 2 and any(kw in word for kw in BUSINESS_KEYWORDS):
|
||||
entities.append(word)
|
||||
|
||||
if entities:
|
||||
return f"{entities[0]} {query}"
|
||||
except ImportError:
|
||||
pass
|
||||
|
||||
return query
|
||||
|
||||
def _detect_image_reference(self, query: str, history: list) -> str:
|
||||
"""检测图片指代查询并重写"""
|
||||
IMAGE_REFERENCE_PATTERNS = [
|
||||
r'这[张些]图片', r'那[张些]图片', r'上面的图片', r'刚才的图片',
|
||||
r'这[张些]图', r'那[张些]图', r'上面的图', r'刚才的图',
|
||||
r'解释一下这[张些]图', r'说明一下这[张些]图',
|
||||
r'这[张些]是什么图', r'图[里内]是什么', r'图片[里内]是什么',
|
||||
]
|
||||
|
||||
is_image_reference = False
|
||||
for pattern in IMAGE_REFERENCE_PATTERNS:
|
||||
if re.search(pattern, query):
|
||||
is_image_reference = True
|
||||
break
|
||||
|
||||
if not is_image_reference:
|
||||
return ""
|
||||
|
||||
last_images = []
|
||||
for msg in reversed(history):
|
||||
if msg.get("role") == "assistant":
|
||||
metadata = msg.get("metadata", {})
|
||||
if isinstance(metadata, dict):
|
||||
images = metadata.get("images", [])
|
||||
if images:
|
||||
for img in images[:5]:
|
||||
if isinstance(img, dict):
|
||||
desc = img.get("description", "")
|
||||
img_type = img.get("type", "图片")
|
||||
if desc:
|
||||
last_images.append(f"{img_type}：{desc}")
|
||||
elif isinstance(img, str):
|
||||
last_images.append(f"图片：{img}")
|
||||
|
||||
if not last_images:
|
||||
content = msg.get("content", "")
|
||||
if "图片" in content or "图表" in content or "图" in content:
|
||||
sentences = content.split("。")
|
||||
for sentence in sentences:
|
||||
if "图片" in sentence or "图表" in sentence:
|
||||
last_images.append(sentence.strip())
|
||||
|
||||
if last_images:
|
||||
break
|
||||
|
||||
if last_images:
|
||||
image_context = " ".join(last_images[:3])
|
||||
question_intent = re.sub(
|
||||
r'这[张些]图片?|那[张些]图片?|上面的图片?|刚才的图片?|解释一下|说明一下',
|
||||
'', query
|
||||
).strip()
|
||||
|
||||
if question_intent:
|
||||
return f"{image_context} {question_intent}"
|
||||
else:
|
||||
return f"详细解释：{image_context}"
|
||||
|
||||
return query
|
||||
|
||||
def _extract_image_context_from_history(self, history: list) -> str:
|
||||
"""从对话历史中提取图片上下文"""
|
||||
if not history:
|
||||
return ""
|
||||
|
||||
for msg in reversed(history):
|
||||
if msg.get("role") == "assistant":
|
||||
metadata = msg.get("metadata", {})
|
||||
images = metadata.get("images", [])
|
||||
content = msg.get("content", "")
|
||||
|
||||
image_descriptions = []
|
||||
|
||||
if images:
|
||||
for i, img in enumerate(images[:5], 1):
|
||||
if isinstance(img, dict):
|
||||
desc = img.get("description", "")
|
||||
img_type = img.get("type", "图片")
|
||||
source = img.get("source", "")
|
||||
page = img.get("page", "")
|
||||
|
||||
img_info = f"图片{i}：{img_type}"
|
||||
if desc:
|
||||
img_info += f"，描述：{desc}"
|
||||
if source:
|
||||
img_info += f"，来源：{source}"
|
||||
if page:
|
||||
img_info += f"，第{page}页"
|
||||
image_descriptions.append(img_info)
|
||||
|
||||
if not image_descriptions:
|
||||
if "图片" in content or "图表" in content:
|
||||
sentences = content.split("。")
|
||||
for sentence in sentences:
|
||||
if "图片" in sentence or "图表" in sentence:
|
||||
image_descriptions.append(sentence.strip())
|
||||
if len(image_descriptions) >= 3:
|
||||
break
|
||||
|
||||
if image_descriptions:
|
||||
return "\n".join(image_descriptions)
|
||||
|
||||
return ""
|
||||
|
||||
def _answer_image_reference(self, enhanced_query: str, history: list) -> str:
|
||||
"""回答图片引用问题"""
|
||||
from core.llm_utils import call_llm
|
||||
|
||||
messages = [
|
||||
{"role": "system", "content": "你是一个专业的助手，请根据提供的图片信息回答用户的问题。"}
|
||||
]
|
||||
|
||||
for h in history[-4:]:
|
||||
if h.get("role") in ["user", "assistant"]:
|
||||
messages.append({"role": h["role"], "content": h.get("content", "")})
|
||||
|
||||
messages.append({"role": "user", "content": enhanced_query})
|
||||
|
||||
try:
|
||||
result = call_llm(
|
||||
self.client, "", MODEL,
|
||||
temperature=0.3,
|
||||
max_tokens=1000,
|
||||
messages=messages
|
||||
)
|
||||
return result or ""
|
||||
except Exception as e:
|
||||
logger.error(f"图片引用回答失败: {e}")
|
||||
return f"抱歉，回答图片问题时出现错误：{str(e)}"
|
||||
|
||||
def _llm_rewrite(self, query: str) -> str:
|
||||
"""LLM 深度重写查询"""
|
||||
from core.llm_utils import call_llm
|
||||
|
||||
prompt = f"""请将以下用户问题改写为更专业、更清晰的表达，保持原意不变。
|
||||
|
||||
原问题：{query}
|
||||
|
||||
改写后的问题："""
|
||||
|
||||
try:
|
||||
rewritten = call_llm(
|
||||
self.client, prompt, MODEL,
|
||||
temperature=0.3,
|
||||
max_tokens=100
|
||||
)
|
||||
return rewritten.strip() if rewritten else query
|
||||
except Exception as e:
|
||||
logger.warning(f"LLM 重写失败: {e}")
|
||||
return query
|
||||
Reference in New Issue
Block a user