134 lines
5.5 KiB
Python
134 lines
5.5 KiB
Python
from src.tools.tool_can_use.base_tool import BaseTool
|
|
from src.chat.utils.utils import get_embedding
|
|
from src.common.database.database_model import Knowledges # Updated import
|
|
from src.common.logger import get_logger
|
|
from typing import Any, Union, List # Added List
|
|
import json # Added for parsing embedding
|
|
import math # Added for cosine similarity
|
|
|
|
logger = get_logger("get_knowledge_tool")
|
|
|
|
|
|
class SearchKnowledgeTool(BaseTool):
|
|
"""从知识库中搜索相关信息的工具"""
|
|
|
|
name = "search_knowledge"
|
|
description = "使用工具从知识库中搜索相关信息"
|
|
parameters = {
|
|
"type": "object",
|
|
"properties": {
|
|
"query": {"type": "string", "description": "搜索查询关键词"},
|
|
"threshold": {"type": "number", "description": "相似度阈值，0.0到1.0之间"},
|
|
},
|
|
"required": ["query"],
|
|
}
|
|
|
|
async def execute(self, function_args: dict[str, Any]) -> dict[str, Any]:
|
|
"""执行知识库搜索
|
|
|
|
Args:
|
|
function_args: 工具参数
|
|
|
|
Returns:
|
|
dict: 工具执行结果
|
|
"""
|
|
query = "" # Initialize query to ensure it's defined in except block
|
|
try:
|
|
query = function_args.get("query")
|
|
threshold = function_args.get("threshold", 0.4)
|
|
|
|
# 调用知识库搜索
|
|
embedding = await get_embedding(query, request_type="info_retrieval")
|
|
if embedding:
|
|
knowledge_info = self.get_info_from_db(embedding, limit=3, threshold=threshold)
|
|
if knowledge_info:
|
|
content = f"你知道这些知识: {knowledge_info}"
|
|
else:
|
|
content = f"你不太了解有关{query}的知识"
|
|
return {"type": "knowledge", "id": query, "content": content}
|
|
return {"type": "info", "id": query, "content": f"无法获取关于'{query}'的嵌入向量，你知识库炸了"}
|
|
except Exception as e:
|
|
logger.error(f"知识库搜索工具执行失败: {str(e)}")
|
|
return {"type": "info", "id": query, "content": f"知识库搜索失败，炸了: {str(e)}"}
|
|
|
|
@staticmethod
|
|
def _cosine_similarity(vec1: List[float], vec2: List[float]) -> float:
|
|
"""计算两个向量之间的余弦相似度"""
|
|
dot_product = sum(p * q for p, q in zip(vec1, vec2, strict=False))
|
|
magnitude1 = math.sqrt(sum(p * p for p in vec1))
|
|
magnitude2 = math.sqrt(sum(q * q for q in vec2))
|
|
if magnitude1 == 0 or magnitude2 == 0:
|
|
return 0.0
|
|
return dot_product / (magnitude1 * magnitude2)
|
|
|
|
@staticmethod
|
|
def get_info_from_db(
|
|
query_embedding: list[float], limit: int = 1, threshold: float = 0.5, return_raw: bool = False
|
|
) -> Union[str, list]:
|
|
"""从数据库中获取相关信息
|
|
|
|
Args:
|
|
query_embedding: 查询的嵌入向量
|
|
limit: 最大返回结果数
|
|
threshold: 相似度阈值
|
|
return_raw: 是否返回原始结果
|
|
|
|
Returns:
|
|
Union[str, list]: 格式化的信息字符串或原始结果列表
|
|
"""
|
|
if not query_embedding:
|
|
return "" if not return_raw else []
|
|
|
|
similar_items = []
|
|
try:
|
|
all_knowledges = Knowledges.select()
|
|
for item in all_knowledges:
|
|
try:
|
|
item_embedding_str = item.embedding
|
|
if not item_embedding_str:
|
|
logger.warning(f"Knowledge item ID {item.id} has empty embedding string.")
|
|
continue
|
|
item_embedding = json.loads(item_embedding_str)
|
|
if not isinstance(item_embedding, list) or not all(
|
|
isinstance(x, (int, float)) for x in item_embedding
|
|
):
|
|
logger.warning(f"Knowledge item ID {item.id} has invalid embedding format after JSON parsing.")
|
|
continue
|
|
except json.JSONDecodeError:
|
|
logger.warning(f"Failed to parse embedding for knowledge item ID {item.id}")
|
|
continue
|
|
except AttributeError:
|
|
logger.warning(f"Knowledge item ID {item.id} missing 'embedding' attribute or it's not a string.")
|
|
continue
|
|
|
|
similarity = SearchKnowledgeTool._cosine_similarity(query_embedding, item_embedding)
|
|
|
|
if similarity >= threshold:
|
|
similar_items.append({"content": item.content, "similarity": similarity, "raw_item": item})
|
|
|
|
# 按相似度降序排序
|
|
similar_items.sort(key=lambda x: x["similarity"], reverse=True)
|
|
|
|
# 应用限制
|
|
results = similar_items[:limit]
|
|
logger.debug(f"知识库查询后，符合条件的结果数量: {len(results)}")
|
|
|
|
except Exception as e:
|
|
logger.error(f"从 Peewee 数据库获取知识信息失败: {str(e)}")
|
|
return "" if not return_raw else []
|
|
|
|
if not results:
|
|
return "" if not return_raw else []
|
|
|
|
if return_raw:
|
|
# Peewee 模型实例不能直接序列化为 JSON，如果需要原始模型，调用者需要处理
|
|
# 这里返回包含内容和相似度的字典列表
|
|
return [{"content": r["content"], "similarity": r["similarity"]} for r in results]
|
|
else:
|
|
# 返回所有找到的内容，用换行分隔
|
|
return "\n".join(str(result["content"]) for result in results)
|
|
|
|
|
|
# 注册工具
|
|
# register_tool(SearchKnowledgeTool)
|