Agent/knowhub/vector_store.py

"""
Milvus Lite 存储封装

单一存储架构，存储完整知识数据 + 向量。
"""

from milvus import default_server
from pymilvus import (
    connections, Collection, FieldSchema,
    CollectionSchema, DataType, utility
)
from typing import List, Dict, Optional
import json
import time


class MilvusStore:
    def __init__(self, data_dir: str = "./milvus_data"):
        """
        初始化 Milvus Lite 存储

        Args:
            data_dir: 数据存储目录
        """
        # 启动内嵌服务器
        default_server.set_base_dir(data_dir)

        # 检查是否已经有 Milvus 实例在运行
        try:
            # 尝试连接到可能已存在的实例
            connections.connect(
                alias="default",
                host='127.0.0.1',
                port=default_server.listen_port,
                timeout=5
            )
            print(f"[Milvus] 连接到已存在的 Milvus 实例 (端口 {default_server.listen_port})")
        except Exception:
            # 没有运行的实例，启动新的
            print(f"[Milvus] 启动新的 Milvus Lite 实例...")
            try:
                default_server.start()
                print(f"[Milvus] Milvus Lite 启动成功 (端口 {default_server.listen_port})")
                # 启动后建立连接
                connections.connect(
                    alias="default",
                    host='127.0.0.1',
                    port=default_server.listen_port,
                    timeout=5
                )
                print(f"[Milvus] 已连接到新启动的实例")
            except Exception as e:
                print(f"[Milvus] 启动失败: {e}")
                # 尝试连接到可能已经在运行的实例
                try:
                    connections.connect(
                        alias="default",
                        host='127.0.0.1',
                        port=default_server.listen_port,
                        timeout=5
                    )
                    print(f"[Milvus] 连接到已存在的实例")
                except Exception as e2:
                    raise RuntimeError(f"无法启动或连接到 Milvus: {e}, {e2}")

        self._init_collection()

    def _init_collection(self):
        """初始化 collection"""
        collection_name = "knowledge"

        if utility.has_collection(collection_name):
            self.collection = Collection(collection_name)
        else:
            # 定义 schema
            fields = [
                FieldSchema(name="id", dtype=DataType.VARCHAR,
                           max_length=100, is_primary=True),
                FieldSchema(name="embedding", dtype=DataType.FLOAT_VECTOR,
                           dim=1536),
                FieldSchema(name="message_id", dtype=DataType.VARCHAR,
                           max_length=100),
                FieldSchema(name="task", dtype=DataType.VARCHAR,
                           max_length=2000),
                FieldSchema(name="content", dtype=DataType.VARCHAR,
                           max_length=50000),
                FieldSchema(name="types", dtype=DataType.ARRAY,
                           element_type=DataType.VARCHAR, max_capacity=20, max_length=50),
                FieldSchema(name="tags", dtype=DataType.JSON),
                FieldSchema(name="scopes", dtype=DataType.ARRAY,
                           element_type=DataType.VARCHAR, max_capacity=20, max_length=100),
                FieldSchema(name="owner", dtype=DataType.VARCHAR,
                           max_length=200),
                FieldSchema(name="resource_ids", dtype=DataType.ARRAY,
                           element_type=DataType.VARCHAR, max_capacity=50, max_length=200),
                FieldSchema(name="source", dtype=DataType.JSON),
                FieldSchema(name="eval", dtype=DataType.JSON),
                FieldSchema(name="created_at", dtype=DataType.INT64),
                FieldSchema(name="updated_at", dtype=DataType.INT64),
            ]

            schema = CollectionSchema(fields, description="KnowHub Knowledge")
            self.collection = Collection(collection_name, schema)

            # 创建向量索引
            index_params = {
                "metric_type": "COSINE",
                "index_type": "HNSW",
                "params": {"M": 16, "efConstruction": 200}
            }
            self.collection.create_index("embedding", index_params)

        self.collection.load()

    def insert(self, knowledge: Dict):
        """
        插入单条知识

        Args:
            knowledge: 知识数据（包含 embedding）
        """
        self.collection.insert([knowledge])
        self.collection.flush()

    def insert_batch(self, knowledge_list: List[Dict]):
        """
        批量插入知识

        Args:
            knowledge_list: 知识列表
        """
        if not knowledge_list:
            return
        self.collection.insert(knowledge_list)
        self.collection.flush()

    def search(self,
               query_embedding: List[float],
               filters: Optional[str] = None,
               limit: int = 10) -> List[Dict]:
        """
        向量检索 + 标量过滤

        Args:
            query_embedding: 查询向量
            filters: 过滤表达式（如: 'owner == "agent"'）
            limit: 返回数量

        Returns:
            知识列表
        """
        search_params = {"metric_type": "COSINE", "params": {"ef": 100}}

        results = self.collection.search(
            data=[query_embedding],
            anns_field="embedding",
            param=search_params,
            limit=limit,
            expr=filters,
            output_fields=["id", "message_id", "task", "content", "types",
                          "tags", "scopes", "owner", "resource_ids",
                          "source", "eval", "created_at", "updated_at"]
        )

        if not results or not results[0]:
            return []

        return [hit.entity.to_dict() for hit in results[0]]

    def query(self, filters: str, limit: int = 100) -> List[Dict]:
        """
        纯标量查询（不使用向量）

        Args:
            filters: 过滤表达式
            limit: 返回数量

        Returns:
            知识列表
        """
        results = self.collection.query(
            expr=filters,
            output_fields=["id", "message_id", "task", "content", "types",
                          "tags", "scopes", "owner", "resource_ids",
                          "source", "eval", "created_at", "updated_at"],
            limit=limit
        )
        return results

    def get_by_id(self, knowledge_id: str) -> Optional[Dict]:
        """
        根据 ID 获取知识

        Args:
            knowledge_id: 知识 ID

        Returns:
            知识数据，不存在返回 None
        """
        results = self.collection.query(
            expr=f'id == "{knowledge_id}"',
            output_fields=["id", "message_id", "task", "content", "types",
                          "tags", "scopes", "owner", "resource_ids",
                          "source", "eval", "created_at", "updated_at"]
        )
        return results[0] if results else None

    def update(self, knowledge_id: str, updates: Dict):
        """
        更新知识（先删除再插入）

        Args:
            knowledge_id: 知识 ID
            updates: 更新字段
        """
        # 1. 查询现有数据
        existing = self.get_by_id(knowledge_id)
        if not existing:
            raise ValueError(f"Knowledge not found: {knowledge_id}")

        # 2. 合并更新
        existing.update(updates)
        existing["updated_at"] = int(time.time())

        # 3. 删除旧数据
        self.delete(knowledge_id)

        # 4. 插入新数据
        self.insert(existing)

    def delete(self, knowledge_id: str):
        """
        删除知识

        Args:
            knowledge_id: 知识 ID
        """
        self.collection.delete(f'id == "{knowledge_id}"')
        self.collection.flush()

    def count(self) -> int:
        """返回知识总数"""
        return self.collection.num_entities

    def drop_collection(self):
        """删除 collection（危险操作）"""
        utility.drop_collection("knowledge")