智能记忆筛选

概述

智能记忆筛选是一个基于多因子综合评分的记忆检索算法，它综合考虑向量相似度、时间遗忘因子和专家评分等多个维度，返回最符合条件的TopK记忆。

核心算法

综合评分公式

composite_score = similarity_score × similarity_weight + 
                 time_score × time_weight + 
                 quality_score × quality_weight

三个关键因子

API 使用方式

1. 底层调用 (memory_manager)

from agents.airport_service.context_engineering.memory_manager import memory_manager

results = await memory_manager.get_smart_filtered_memories(
    query="机场餐厅推荐",
    user_id="user123",
    agent_name="机场知识问答子智能体",
    limit=10,
    similarity_weight=0.5,    # 相似度权重
    time_weight=0.2,          # 时间权重  
    quality_weight=0.3,       # 质量权重
    min_quality_score=0.7,    # 最低质量门槛
    time_decay_days=30        # 时间衰减周期(天)
)

2. 智能体集成层调用

from agents.airport_service.context_engineering.agent_memory import AgentMemoryMixin

smart_memories = await AgentMemoryMixin.retrieve_smart_filtered_memories(
    user_id="user123",
    current_query="机场餐厅推荐",
    agent_name="机场知识问答子智能体",
    limit=5,
    similarity_weight=0.6,
    time_weight=0.2,
    quality_weight=0.2
)

3. HTTP API 调用

POST /memory/v1/conversations/smart-filter
Content-Type: application/json

{
    "query": "机场餐厅推荐",
    "user_id": "user123",
    "agent_name": "机场知识问答子智能体",
    "limit": 10,
    "similarity_weight": 0.5,
    "time_weight": 0.2,
    "quality_weight": 0.3,
    "min_quality_score": 0.7,
    "time_decay_days": 30
}

参数配置指南

权重配置策略

similarity_weight=0.3
time_weight=0.2  
quality_weight=0.5  # 高质量权重

similarity_weight=0.3
time_weight=0.5     # 高时间权重
quality_weight=0.2

similarity_weight=0.7  # 高相似度权重
time_weight=0.15
quality_weight=0.15

similarity_weight=0.4
time_weight=0.3
quality_weight=0.3

时间衰减周期设置

质量门槛设置

返回数据结构

{
    "memory_id": "mem_123",
    "user_id": "user123", 
    "query": "用户查询",
    "response": "智能体回复",
    "agent_name": "智能体名称",
    "expert_verified": true,
    "quality_score": 0.95,
    "created_at": "2024-01-01T12:00:00",
    
    // 评分详情
    "similarity_score": 0.85,
    "time_score": 0.92,
    "composite_score": 0.89,
    "expert_corrected": false,
    
    "score_breakdown": {
        "similarity": 0.85,
        "time_factor": 0.92,
        "quality": 0.95,
        "weights": {
            "similarity": 0.5,
            "time": 0.2,
            "quality": 0.3
        }
    }
}

最佳实践

使用示例

运行演示程序查看完整示例：

python examples/smart_memory_filter_demo.py

实际应用示例

# 重视准确性和时效性
memories = await memory_manager.get_smart_filtered_memories(
    query="液体登机规定",
    user_id="user001",
    agent_name="安检助手",
    similarity_weight=0.4,
    time_weight=0.3,
    quality_weight=0.3,
    min_quality_score=0.8,
    time_decay_days=30
)

# 重视时效性
memories = await memory_manager.get_smart_filtered_memories(
    query="航班延误通知",
    user_id="user002",
    agent_name="航班助手",
    similarity_weight=0.3,
    time_weight=0.5,
    quality_weight=0.2,
    time_decay_days=7
)

常见问题

扩展功能

自定义评分因子

如需添加更多评分维度(如用户反馈、点击率等)，可以扩展get_smart_filtered_memories方法：

# 在memory_manager.py中扩展
custom_score = calculate_custom_factor(result)
composite_score = (
    similarity_score * similarity_weight +
    time_score * time_weight +
    quality_score * quality_weight +
    custom_score * custom_weight
)

这样就可以根据业务需求灵活扩展评分模型。

多维度评分可视化