|
@@ -0,0 +1,765 @@
|
|
|
|
|
+import asyncio
|
|
|
|
|
+import json
|
|
|
|
|
+import os
|
|
|
|
|
+import argparse
|
|
|
|
|
+from datetime import datetime
|
|
|
|
|
+
|
|
|
|
|
+from agents import Agent, Runner
|
|
|
|
|
+from lib.my_trace import set_trace
|
|
|
|
|
+from typing import Literal
|
|
|
|
|
+from pydantic import BaseModel, Field
|
|
|
|
|
+
|
|
|
|
|
+from lib.utils import read_file_as_string
|
|
|
|
|
+from script.search_recommendations.xiaohongshu_search_recommendations import XiaohongshuSearchRecommendations
|
|
|
|
|
+
|
|
|
|
|
+
|
|
|
|
|
+class RunContext(BaseModel):
|
|
|
|
|
+ version: str = Field(..., description="当前运行的脚本版本(文件名)")
|
|
|
|
|
+ input_files: dict[str, str] = Field(..., description="输入文件路径映射")
|
|
|
|
|
+ q_with_context: str
|
|
|
|
|
+ q_context: str
|
|
|
|
|
+ q: str
|
|
|
|
|
+ log_url: str
|
|
|
|
|
+ log_dir: str
|
|
|
|
|
+ question_annotation: str | None = Field(default=None, description="问题的标注结果")
|
|
|
|
|
+ operations_history: list[dict] = Field(default_factory=list, description="记录所有操作的历史")
|
|
|
|
|
+ final_output: str | None = Field(default=None, description="最终输出结果")
|
|
|
|
|
+
|
|
|
|
|
+
|
|
|
|
|
+# ============================================================================
|
|
|
|
|
+# Agent 1: 问题标注专家
|
|
|
|
|
+# ============================================================================
|
|
|
|
|
+question_annotation_instructions = """
|
|
|
|
|
+你是搜索需求分析专家。给定问题(含需求背景),在原文上标注三层:本质、硬性、软性。
|
|
|
|
|
+
|
|
|
|
|
+## 判断标准
|
|
|
|
|
+
|
|
|
|
|
+**[本质]** - 问题的核心意图
|
|
|
|
|
+- 如何获取、教程、推荐、作品、测评等
|
|
|
|
|
+
|
|
|
|
|
+**[硬]** - 客观事实性约束(可明确验证、非主观判断)
|
|
|
|
|
+- 能明确区分类别的:地域、时间、对象、工具、操作类型
|
|
|
|
|
+- 特征:改变后得到完全不同类别的结果
|
|
|
|
|
+
|
|
|
|
|
+**[软]** - 主观判断性修饰(因人而异、程度性的)
|
|
|
|
|
+- 需要主观评价的:质量、速度、美观、特色、程度
|
|
|
|
|
+- 特征:改变后仍是同类结果,只是满足程度不同
|
|
|
|
|
+
|
|
|
|
|
+## 输出格式
|
|
|
|
|
+
|
|
|
|
|
+词语[本质-描述]、词语[硬-描述]、词语[软-描述]
|
|
|
|
|
+
|
|
|
|
|
+## 注意
|
|
|
|
|
+- 只输出标注后的字符串
|
|
|
|
|
+- 结合需求背景判断意图
|
|
|
|
|
+""".strip()
|
|
|
|
|
+
|
|
|
|
|
+question_annotator = Agent[None](
|
|
|
|
|
+ name="问题标注专家",
|
|
|
|
|
+ instructions=question_annotation_instructions,
|
|
|
|
|
+)
|
|
|
|
|
+
|
|
|
|
|
+
|
|
|
|
|
+# ============================================================================
|
|
|
|
|
+# Agent 2: 评估专家
|
|
|
|
|
+# ============================================================================
|
|
|
|
|
+eval_instructions = """
|
|
|
|
|
+你是搜索query评估专家。给定原始问题标注(三层)和推荐query,评估三个分数。
|
|
|
|
|
+
|
|
|
|
|
+## 评估目标
|
|
|
|
|
+
|
|
|
|
|
+用这个推荐query搜索,能否找到满足原始需求的内容?
|
|
|
|
|
+
|
|
|
|
|
+## 三层评分
|
|
|
|
|
+
|
|
|
|
|
+### 1. essence_score(本质/意图)= 0 或 1
|
|
|
|
|
+
|
|
|
|
|
+推荐query的本质/意图是否与原问题一致?
|
|
|
|
|
+
|
|
|
|
|
+**原问题标注中的[本质-XXX]对应推荐词要求:**
|
|
|
|
|
+- 找方法/如何获取 → 推荐词**必须明确包含方法/途径类词汇**
|
|
|
|
|
+ - ✅ "...网站推荐"、"如何获取..."、"...获取途径"、"...方法"
|
|
|
|
|
+ - ❌ "...下载"、"...素材"(直接找内容,不是找方法)
|
|
|
|
|
+- 教程/学习 → 推荐词应该是教程/教学
|
|
|
|
|
+- 作品/欣赏 → 推荐词应该是作品展示
|
|
|
|
|
+- 工具/推荐 → 推荐词应该是工具推荐
|
|
|
|
|
+
|
|
|
|
|
+**评分原则:**
|
|
|
|
|
+- 1 = 本质一致,推荐词**明确表达**相同意图
|
|
|
|
|
+- 0 = 本质改变或**不够明确**(宁可严格,不可放松)
|
|
|
|
|
+
|
|
|
|
|
+### 2. hard_score(硬性约束)= 0 或 1
|
|
|
|
|
+
|
|
|
|
|
+在本质一致的前提下,是否满足所有硬性约束?
|
|
|
|
|
+
|
|
|
|
|
+**原问题标注中的[硬-XXX]:**地域、时间、对象、质量、工具等
|
|
|
|
|
+
|
|
|
|
|
+**评分:**
|
|
|
|
|
+- 1 = 所有硬性约束都满足
|
|
|
|
|
+- 0 = 任一硬性约束不满足
|
|
|
|
|
+
|
|
|
|
|
+### 3. soft_score(软性修饰)= 0-1
|
|
|
|
|
+
|
|
|
|
|
+软性修饰词保留了多少?
|
|
|
|
|
+
|
|
|
|
|
+**评分参考:**
|
|
|
|
|
+- 1.0 = 完整保留
|
|
|
|
|
+- 0.7-0.9 = 保留核心
|
|
|
|
|
+- 0.4-0.6 = 部分丢失
|
|
|
|
|
+- 0-0.3 = 大量丢失
|
|
|
|
|
+
|
|
|
|
|
+## 注意
|
|
|
|
|
+
|
|
|
|
|
+- essence=0 直接拒绝,不管hard/soft多高
|
|
|
|
|
+- essence=1, hard=0 也要拒绝
|
|
|
|
|
+- essence=1, hard=1 才看soft_score
|
|
|
|
|
+""".strip()
|
|
|
|
|
+
|
|
|
|
|
+class EvaluationFeedback(BaseModel):
|
|
|
|
|
+ """评估反馈模型 - 三层评分"""
|
|
|
|
|
+ essence_score: Literal[0, 1] = Field(..., description="本质/意图匹配度,0或1")
|
|
|
|
|
+ hard_score: Literal[0, 1] = Field(..., description="硬性约束匹配度,0或1")
|
|
|
|
|
+ soft_score: float = Field(..., description="软性修饰完整度,0-1")
|
|
|
|
|
+ reason: str = Field(..., description="评估理由")
|
|
|
|
|
+
|
|
|
|
|
+evaluator = Agent[None](
|
|
|
|
|
+ name="评估专家",
|
|
|
|
|
+ instructions=eval_instructions,
|
|
|
|
|
+ output_type=EvaluationFeedback,
|
|
|
|
|
+)
|
|
|
|
|
+
|
|
|
|
|
+
|
|
|
|
|
+# ============================================================================
|
|
|
|
|
+# Agent 3: 修改策略生成专家
|
|
|
|
|
+# ============================================================================
|
|
|
|
|
+strategy_instructions = """
|
|
|
|
|
+你是query修改策略专家。**模拟人在搜索引擎中的真实搜索行为**,基于反馈动态调整query。
|
|
|
|
|
+
|
|
|
|
|
+## 核心思路:搜索是探索过程,不是直达过程
|
|
|
|
|
+
|
|
|
|
|
+**关键认知:**
|
|
|
|
|
+1. **中间query不需要满足原始需求** - 它是探索工具,可以偏离原需求
|
|
|
|
|
+2. **推荐词是最宝贵的反馈信号** - 告诉你系统理解成什么了,有什么内容
|
|
|
|
|
+3. **每一步query都有明确的探索目的** - 不是盲目改词,而是试探和引导
|
|
|
|
|
+4. **最终目标:找到满足需求的推荐词** - 不是让query本身满足需求
|
|
|
|
|
+
|
|
|
|
|
+## 人的真实搜索过程
|
|
|
|
|
+
|
|
|
|
|
+**搜索的本质**:通过多步探索,利用推荐词作为桥梁,逐步引导系统
|
|
|
|
|
+
|
|
|
|
|
+**典型模式**:
|
|
|
|
|
+
|
|
|
|
|
+第1步:直接尝试
|
|
|
|
|
+- 目的:看系统能否直接理解
|
|
|
|
|
+- 结果:空列表或essence=0
|
|
|
|
|
+- essence=0的推荐词:告诉你系统理解成什么了
|
|
|
|
|
+
|
|
|
|
|
+第2步:降低要求,简化query
|
|
|
|
|
+- 目的:让系统有响应,看它在基础层面有什么
|
|
|
|
|
+- 推荐词虽然essence=0,但揭示了系统在某个主题有内容
|
|
|
|
|
+- **关键**:选一个最有潜力的推荐词
|
|
|
|
|
+
|
|
|
|
|
+第3步:基于推荐词,往目标方向引导
|
|
|
|
|
+- 目的:利用推荐词作为桥梁,加上目标方向的词
|
|
|
|
|
+- 推荐词还是essence=0,但主题在变化(接近目标)
|
|
|
|
|
+- **渐进式**:不求一步到位,每步都有进展
|
|
|
|
|
+
|
|
|
|
|
+第4步:继续引导或换角度
|
|
|
|
|
+- 如果推荐词主题不变 → 换角度
|
|
|
|
|
+- 如果推荐词主题在接近 → 继续引导
|
|
|
|
|
+
|
|
|
|
|
+最终:找到essence=1的推荐词
|
|
|
|
|
+
|
|
|
|
|
+**关键原则**:
|
|
|
|
|
+1. essence_score是评估推荐词的,不是评估中间query的
|
|
|
|
|
+2. essence=0的推荐词也有价值,它揭示了系统的理解方向
|
|
|
|
|
+3. 每一步都有明确的探索目的,看目的是否达成
|
|
|
|
|
+4. 通过推荐词的主题变化,判断是否在接近目标
|
|
|
|
|
+
|
|
|
|
|
+## 输入信息
|
|
|
|
|
+- 原始问题标注(三层):本质、硬性约束、软性修饰
|
|
|
|
|
+- 历史尝试记录:所有轮次的query、推荐词、评估结果
|
|
|
|
|
+- 当前query和推荐词评估
|
|
|
|
|
+
|
|
|
|
|
+## 分析步骤
|
|
|
|
|
+
|
|
|
|
|
+### 第一步:理解当前推荐词的信号
|
|
|
|
|
+**核心问题:推荐词告诉我什么信息?**
|
|
|
|
|
+
|
|
|
|
|
+**重要提醒:essence_score是评估推荐词是否满足原始需求的最终目标**
|
|
|
|
|
+- essence_score=1: 推荐词满足原需求的本质
|
|
|
|
|
+- essence_score=0: 推荐词不满足原需求的本质
|
|
|
|
|
+- **但中间query的目的可能不是满足原需求**,所以essence_score只是参考
|
|
|
|
|
+
|
|
|
|
|
+1. **系统理解层面**(看推荐词的主题):
|
|
|
|
|
+ - 空列表 → 系统完全不理解当前query
|
|
|
|
|
+ - 有推荐词 → 系统理解成了什么主题?
|
|
|
|
|
+ - 旅游?教程?素材?工具?品种介绍?
|
|
|
|
|
+ - 这些主题是否有助于往目标方向引导?
|
|
|
|
|
+
|
|
|
|
|
+2. **内容可用性层面**(看推荐词的价值):
|
|
|
|
|
+ - **即使推荐词essence=0,也可能是很好的探索起点**
|
|
|
|
|
+ - 例如:推荐词"川西旅游攻略"虽然essence=0,但揭示了系统认识"川西"
|
|
|
|
|
+ - 哪些推荐词最有潜力作为下一步的桥梁?
|
|
|
|
|
+
|
|
|
|
|
+3. **探索目的验证**:
|
|
|
|
|
+ - 当前query的探索目的是什么?达到了吗?
|
|
|
|
|
+ - 例如:目的是"看系统对川西有什么" → 达到了(有推荐词)
|
|
|
|
|
+ - 下一步要验证/探索什么?
|
|
|
|
|
+
|
|
|
|
|
+### 第二步:回顾历史,识别规律
|
|
|
|
|
+
|
|
|
|
|
+**什么是真正的死胡同?**
|
|
|
|
|
+- ❌ 不是essence=0就是死胡同
|
|
|
|
|
+- ✅ 死胡同:连续多次尝试,**推荐词的主题完全不变**
|
|
|
|
|
+ - 例如:连续3轮都是"取名/品种介绍"
|
|
|
|
|
+ - 例如:连续3轮都是"旅游攻略"
|
|
|
|
|
+
|
|
|
|
|
+**什么是有进展?**
|
|
|
|
|
+- ✅ 推荐词的主题在变化(旅游→摄影→作品→素材)
|
|
|
|
|
+- ✅ 即使essence=0,也说明在接近目标
|
|
|
|
|
+- ✅ 说明系统的理解在被引导
|
|
|
|
|
+
|
|
|
|
|
+**避免同义词打转**:
|
|
|
|
|
+- ❌ "素材"→"照片"→"图片"→"资源"
|
|
|
|
|
+- 这些同义词不会改变系统理解,纯粹浪费轮次
|
|
|
|
|
+- ✅ 应该从推荐词中选一个作为桥梁,或换完全不同的角度
|
|
|
|
|
+
|
|
|
|
|
+### 第三步:选择策略类型(带着明确的探索目的)
|
|
|
|
|
+
|
|
|
|
|
+**refine_current(微调当前query)**
|
|
|
|
|
+- 适用:推荐词方向对了,需要微调让它更精确
|
|
|
|
|
+- 探索目的:在正确方向上精细化
|
|
|
|
|
+- 动作:加词/减词/换词/调整顺序
|
|
|
|
|
+
|
|
|
|
|
+**use_recommendation(选推荐词作为新起点)** ⭐ 最重要策略
|
|
|
|
|
+- 适用:推荐词虽然essence=0,但**揭示了系统在这个方向有内容**
|
|
|
|
|
+- 探索目的:利用推荐词这个客观信号,引导系统往目标方向
|
|
|
|
|
+- **核心思维**:推荐词是系统给你的提示,告诉你"我有这个"
|
|
|
|
|
+- 动作:
|
|
|
|
|
+ - 选一个最有潜力的推荐词作为base_query
|
|
|
|
|
+ - 在它基础上加目标方向的词
|
|
|
|
|
+ - **这个新query可能不满足原需求,但目的是探索和引导**
|
|
|
|
|
+
|
|
|
|
|
+**change_approach(换完全不同的角度)**
|
|
|
|
|
+- 适用:当前方向是死路(多次尝试推荐词主题不变)
|
|
|
|
|
+- 探索目的:跳出当前框架,从另一个角度切入
|
|
|
|
|
+- 动作:换一种完全不同的表述方式
|
|
|
|
|
+
|
|
|
|
|
+**relax_constraints(放宽约束)**
|
|
|
|
|
+- 适用:query太复杂,系统不理解(返回空列表)
|
|
|
|
|
+- 探索目的:先让系统有响应,看它在最基础层面有什么
|
|
|
|
|
+- 动作:去掉限定词,保留核心概念
|
|
|
|
|
+
|
|
|
|
|
+## 输出要求
|
|
|
|
|
+
|
|
|
|
|
+### 1. reasoning(推理过程)
|
|
|
|
|
+必须包含三部分,**核心是目标导向**:
|
|
|
|
|
+
|
|
|
|
|
+- **当前位置评估**(相对于最终目标):
|
|
|
|
|
+ - 当前推荐词反映了系统在什么主题空间?
|
|
|
|
|
+ - 离原始需求多远?(完全偏离/部分相关/已经接近)
|
|
|
|
|
+ - **不要只看essence_score**:
|
|
|
|
|
+ - essence_score=0不代表没用,可能是必经之路
|
|
|
|
|
+ - 关键看推荐词能否帮助靠近目标
|
|
|
|
|
+
|
|
|
|
|
+- **历史进展判断**:
|
|
|
|
|
+ - 推荐词主题是否在变化?(变化=进展,不变=死胡同)
|
|
|
|
|
+ - 在靠近目标,还是原地打转?
|
|
|
|
|
+ - 哪些尝试让系统理解发生了改变?
|
|
|
|
|
+
|
|
|
|
|
+- **下一步探索规划**:
|
|
|
|
|
+ - **这一步的作用**(必须明确!):
|
|
|
|
|
+ * 离目标远?→ 先让系统理解某个关键词/概念
|
|
|
|
|
+ * 部分相关?→ 选推荐词作为桥梁,引导往目标方向
|
|
|
|
|
+ * 已经接近?→ 微调细节,精确匹配
|
|
|
|
|
+ - 为什么选这个base_query?
|
|
|
|
|
+ - 这个修改如何让我们靠近目标?
|
|
|
|
|
+ - **重要**:中间query不需要满足原需求,只要能靠近目标
|
|
|
|
|
+
|
|
|
|
|
+### 2. strategy_type
|
|
|
|
|
+从4种策略中选择:refine_current, use_recommendation, change_approach, relax_constraints
|
|
|
|
|
+
|
|
|
|
|
+### 3. base_query
|
|
|
|
|
+**关键**:可以选择历史中的query,也可以选择历史推荐词
|
|
|
|
|
+- 如果选历史query:base_query_source = "history_query"
|
|
|
|
|
+- 如果选历史推荐词:base_query_source = "history_recommendation"
|
|
|
|
|
+
|
|
|
|
|
+### 4. base_query_source
|
|
|
|
|
+说明base_query的来源
|
|
|
|
|
+
|
|
|
|
|
+### 5. modification_action
|
|
|
|
|
+**重要:一次只做一个核心动作**
|
|
|
|
|
+- 不要列举多个动作
|
|
|
|
|
+- 只描述最核心的那一个修改
|
|
|
|
|
+- 例如:"选择推荐词'川西旅游'作为新起点"
|
|
|
|
|
+- 例如:"去掉'如何获取'改为直接搜内容"
|
|
|
|
|
+- 例如:"加上'AI生成'转向生成方向"
|
|
|
|
|
+
|
|
|
|
|
+### 6. new_query
|
|
|
|
|
+最终的新query
|
|
|
|
|
+
|
|
|
|
|
+## 重要原则
|
|
|
|
|
+
|
|
|
|
|
+1. **推荐词是最宝贵的反馈** - 充分利用推荐词这个客观信号
|
|
|
|
|
+ - 即使essence=0的推荐词,也揭示了系统在这个方向有什么
|
|
|
|
|
+ - **优先考虑use_recommendation策略** - 选一个推荐词作为起点
|
|
|
|
|
+
|
|
|
|
|
+2. **中间query可以偏离原需求** - 每一步都有明确的探索目的
|
|
|
|
|
+ - 不要纠结"这个query不满足原需求"
|
|
|
|
|
+ - 关键是:这个query能不能帮你往正确方向引导系统
|
|
|
|
|
+
|
|
|
|
|
+3. **识别死胡同,及时换方向**
|
|
|
|
|
+ - 如果多次尝试推荐词主题不变 → 换方向
|
|
|
|
|
+ - 如果推荐词越来越偏 → 回退到之前的某个好的起点
|
|
|
|
|
+
|
|
|
|
|
+4. **保持推理简洁** - 抓住关键信息
|
|
|
|
|
+ - 明确说出探索目的
|
|
|
|
|
+ - 不要重复啰嗦
|
|
|
|
|
+""".strip()
|
|
|
|
|
+
|
|
|
|
|
+class ModificationStrategy(BaseModel):
|
|
|
|
|
+ """修改策略模型 - 模拟人的搜索调整过程"""
|
|
|
|
|
+ reasoning: str = Field(..., description="推理过程:1)当前推荐词分析:系统理解成什么了?2)历史尝试总结:哪些方向有效/无效?3)下一步策略:为什么这样调整?")
|
|
|
|
|
+
|
|
|
|
|
+ strategy_type: Literal[
|
|
|
|
|
+ "refine_current", # 微调当前query(加词/减词/换词/换顺序)
|
|
|
|
|
+ "use_recommendation", # 选择推荐词作为新起点,在它基础上修改
|
|
|
|
|
+ "change_approach", # 换完全不同的表述角度
|
|
|
|
|
+ "relax_constraints" # 放宽约束,去掉部分限定词
|
|
|
|
|
+ ] = Field(..., description="策略类型")
|
|
|
|
|
+
|
|
|
|
|
+ base_query: str = Field(..., description="基础query,可以是:1)历史中的query 2)历史推荐词中的某一个")
|
|
|
|
|
+ base_query_source: Literal["history_query", "history_recommendation"] = Field(..., description="base_query的来源")
|
|
|
|
|
+
|
|
|
|
|
+ modification_action: str = Field(..., description="核心修改动作(只一个),如:'选择推荐词作为新起点' 或 '去掉方法类词改为直接搜内容' 或 '加上AI生成转向生成方向'")
|
|
|
|
|
+
|
|
|
|
|
+ new_query: str = Field(..., description="修改后的新query")
|
|
|
|
|
+
|
|
|
|
|
+strategy_generator = Agent[None](
|
|
|
|
|
+ name="策略生成专家",
|
|
|
|
|
+ instructions=strategy_instructions,
|
|
|
|
|
+ output_type=ModificationStrategy,
|
|
|
|
|
+)
|
|
|
|
|
+
|
|
|
|
|
+
|
|
|
|
|
+# ============================================================================
|
|
|
|
|
+# 核心函数
|
|
|
|
|
+# ============================================================================
|
|
|
|
|
+
|
|
|
|
|
+async def annotate_question(q_with_context: str) -> str:
|
|
|
|
|
+ """标注问题(三层)"""
|
|
|
|
|
+ print("\n正在标注问题...")
|
|
|
|
|
+ result = await Runner.run(question_annotator, q_with_context)
|
|
|
|
|
+ annotation = str(result.final_output)
|
|
|
|
|
+ print(f"问题标注完成:{annotation}")
|
|
|
|
|
+ return annotation
|
|
|
|
|
+
|
|
|
|
|
+
|
|
|
|
|
+async def get_suggestions_with_eval(query: str, annotation: str, context: RunContext) -> list[dict]:
|
|
|
|
|
+ """获取推荐词并评估"""
|
|
|
|
|
+ print(f"\n正在获取推荐词:{query}")
|
|
|
|
|
+
|
|
|
|
|
+ # 1. 调用小红书API
|
|
|
|
|
+ xiaohongshu_api = XiaohongshuSearchRecommendations()
|
|
|
|
|
+ query_suggestions = xiaohongshu_api.get_recommendations(keyword=query)
|
|
|
|
|
+ print(f"获取到 {len(query_suggestions) if query_suggestions else 0} 个推荐词:{query_suggestions}")
|
|
|
|
|
+
|
|
|
|
|
+ if not query_suggestions:
|
|
|
|
|
+ # 记录到历史
|
|
|
|
|
+ context.operations_history.append({
|
|
|
|
|
+ "operation_type": "get_query_suggestions",
|
|
|
|
|
+ "timestamp": datetime.now().isoformat(),
|
|
|
|
|
+ "query": query,
|
|
|
|
|
+ "suggestions": [],
|
|
|
|
|
+ "evaluations": "未返回任何推荐词",
|
|
|
|
|
+ })
|
|
|
|
|
+ return []
|
|
|
|
|
+
|
|
|
|
|
+ # 2. 并发评估所有推荐词
|
|
|
|
|
+ async def evaluate_single_query(q_sug: str):
|
|
|
|
|
+ eval_input = f"""
|
|
|
|
|
+<需求背景>
|
|
|
|
|
+{context.q_context if context.q_context else "无"}
|
|
|
|
|
+</需求背景>
|
|
|
|
|
+
|
|
|
|
|
+<原始问题>
|
|
|
|
|
+{context.q}
|
|
|
|
|
+</原始问题>
|
|
|
|
|
+
|
|
|
|
|
+<原始问题标注(三层)>
|
|
|
|
|
+{annotation}
|
|
|
|
|
+</原始问题标注(三层)>
|
|
|
|
|
+
|
|
|
|
|
+<待评估的推荐query>
|
|
|
|
|
+{q_sug}
|
|
|
|
|
+</待评估的推荐query>
|
|
|
|
|
+
|
|
|
|
|
+请评估该推荐query的三个分数:
|
|
|
|
|
+1. essence_score: 本质/意图是否一致(0或1)
|
|
|
|
|
+2. hard_score: 硬性约束是否满足(0或1)
|
|
|
|
|
+3. soft_score: 软性修饰保留程度(0-1)
|
|
|
|
|
+4. reason: 详细的评估理由
|
|
|
|
|
+"""
|
|
|
|
|
+ evaluator_result = await Runner.run(evaluator, eval_input)
|
|
|
|
|
+ result: EvaluationFeedback = evaluator_result.final_output
|
|
|
|
|
+ return {
|
|
|
|
|
+ "query": q_sug,
|
|
|
|
|
+ "essence_score": result.essence_score,
|
|
|
|
|
+ "hard_score": result.hard_score,
|
|
|
|
|
+ "soft_score": result.soft_score,
|
|
|
|
|
+ "reason": result.reason,
|
|
|
|
|
+ }
|
|
|
|
|
+
|
|
|
|
|
+ evaluations = await asyncio.gather(*[evaluate_single_query(q_sug) for q_sug in query_suggestions])
|
|
|
|
|
+
|
|
|
|
|
+ # 3. 记录到历史
|
|
|
|
|
+ context.operations_history.append({
|
|
|
|
|
+ "operation_type": "get_query_suggestions",
|
|
|
|
|
+ "timestamp": datetime.now().isoformat(),
|
|
|
|
|
+ "query": query,
|
|
|
|
|
+ "suggestions": query_suggestions,
|
|
|
|
|
+ "evaluations": evaluations,
|
|
|
|
|
+ })
|
|
|
|
|
+
|
|
|
|
|
+ return evaluations
|
|
|
|
|
+
|
|
|
|
|
+
|
|
|
|
|
+async def generate_modification_strategy(
|
|
|
|
|
+ annotation: str,
|
|
|
|
|
+ context: RunContext
|
|
|
|
|
+) -> ModificationStrategy:
|
|
|
|
|
+ """生成修改策略"""
|
|
|
|
|
+ print("\n正在生成修改策略...")
|
|
|
|
|
+
|
|
|
|
|
+ # 整理历史尝试记录 - 完整保留推荐词和评估结果
|
|
|
|
|
+ history_records = []
|
|
|
|
|
+ round_num = 0
|
|
|
|
|
+
|
|
|
|
|
+ for op in context.operations_history:
|
|
|
|
|
+ if op["operation_type"] == "get_query_suggestions":
|
|
|
|
|
+ round_num += 1
|
|
|
|
|
+ record = {
|
|
|
|
|
+ "round": round_num,
|
|
|
|
|
+ "query": op["query"],
|
|
|
|
|
+ "suggestions": op["suggestions"],
|
|
|
|
|
+ "evaluations": op["evaluations"]
|
|
|
|
|
+ }
|
|
|
|
|
+ history_records.append(record)
|
|
|
|
|
+ elif op["operation_type"] == "modify_query":
|
|
|
|
|
+ # 修改操作也记录,但不增加轮数
|
|
|
|
|
+ history_records.append({
|
|
|
|
|
+ "operation": "modify_query",
|
|
|
|
|
+ "strategy_type": op.get("strategy_type", op.get("modification_type")), # 兼容旧字段
|
|
|
|
|
+ "base_query": op.get("base_query"),
|
|
|
|
|
+ "base_query_source": op.get("base_query_source"),
|
|
|
|
|
+ "modification_action": op.get("modification_action", op.get("modification_actions", [])), # 兼容旧版本
|
|
|
|
|
+ "original_query": op["original_query"],
|
|
|
|
|
+ "new_query": op["new_query"],
|
|
|
|
|
+ "reasoning": op["reasoning"]
|
|
|
|
|
+ })
|
|
|
|
|
+
|
|
|
|
|
+ # 格式化历史记录为JSON
|
|
|
|
|
+ history_json = json.dumps(history_records, ensure_ascii=False, indent=2)
|
|
|
|
|
+
|
|
|
|
|
+ strategy_input = f"""
|
|
|
|
|
+<需求背景>
|
|
|
|
|
+{context.q_context if context.q_context else "无"}
|
|
|
|
|
+</需求背景>
|
|
|
|
|
+
|
|
|
|
|
+<原始问题>
|
|
|
|
|
+{context.q}
|
|
|
|
|
+</原始问题>
|
|
|
|
|
+
|
|
|
|
|
+<原始问题标注(三层)>
|
|
|
|
|
+{annotation}
|
|
|
|
|
+</原始问题标注(三层)>
|
|
|
|
|
+
|
|
|
|
|
+<历史尝试记录(完整)>
|
|
|
|
|
+{history_json}
|
|
|
|
|
+</历史尝试记录(完整)>
|
|
|
|
|
+
|
|
|
|
|
+请基于所有历史尝试,生成下一步的query修改策略。
|
|
|
|
|
+
|
|
|
|
|
+**说明**:历史记录中最后一条就是当前轮的query和推荐词评估结果。
|
|
|
|
|
+
|
|
|
|
|
+**核心思路**:每一步都要明确 "当前在哪 → 离目标多远 → 下一步做什么能靠近目标"
|
|
|
|
|
+
|
|
|
|
|
+重点分析:
|
|
|
|
|
+
|
|
|
|
|
+1. **评估当前位置**(相对于最终目标):
|
|
|
|
|
+ - 当前推荐词反映了什么?系统在哪个主题空间?
|
|
|
|
|
+ - 离原始需求的距离:完全偏离?部分相关?已经接近?
|
|
|
|
|
+ - **不要只看essence_score**:essence=0不代表没用,可能是通往目标的必经之路
|
|
|
|
|
+
|
|
|
|
|
+2. **判断历史进展**:
|
|
|
|
|
+ - 推荐词主题是否在变化?(变化=有进展,不变=死胡同)
|
|
|
|
|
+ - 是在靠近目标,还是在原地打转?
|
|
|
|
|
+ - 哪个方向的query让系统理解发生了改变?
|
|
|
|
|
+
|
|
|
|
|
+3. **规划下一步探索**:
|
|
|
|
|
+ - **这一步query的作用是什么**?(必须明确!)
|
|
|
|
|
+ * 如果离目标很远:需要先让系统理解某个关键词/概念
|
|
|
|
|
+ * 如果部分相关:选一个推荐词作为桥梁,在它基础上引导
|
|
|
|
|
+ * 如果已经接近:微调细节,精确匹配需求
|
|
|
|
|
+ - **记住**:中间query不需要满足原需求,只要能让我们往目标靠近
|
|
|
|
|
+"""
|
|
|
|
|
+ result = await Runner.run(strategy_generator, strategy_input)
|
|
|
|
|
+ strategy: ModificationStrategy = result.final_output
|
|
|
|
|
+ return strategy
|
|
|
|
|
+
|
|
|
|
|
+
|
|
|
|
|
+def find_qualified_queries(evaluations: list[dict], min_soft_score: float = 0.7) -> list[dict]:
|
|
|
|
|
+ """查找所有合格的query,按soft_score降序排列"""
|
|
|
|
|
+ qualified = [
|
|
|
|
|
+ e for e in evaluations
|
|
|
|
|
+ if e['essence_score'] == 1
|
|
|
|
|
+ and e['hard_score'] == 1
|
|
|
|
|
+ and e['soft_score'] >= min_soft_score
|
|
|
|
|
+ ]
|
|
|
|
|
+ # 按soft_score降序排列
|
|
|
|
|
+ return sorted(qualified, key=lambda x: x['soft_score'], reverse=True)
|
|
|
|
|
+
|
|
|
|
|
+
|
|
|
|
|
+# ============================================================================
|
|
|
|
|
+# 主流程(代码控制)
|
|
|
|
|
+# ============================================================================
|
|
|
|
|
+
|
|
|
|
|
+async def optimize_query(context: RunContext, max_rounds: int = 20) -> dict:
|
|
|
|
|
+ """
|
|
|
|
|
+ 主优化流程 - 由代码控制
|
|
|
|
|
+
|
|
|
|
|
+ Args:
|
|
|
|
|
+ context: 运行上下文
|
|
|
|
|
+ max_rounds: 最大迭代轮数,默认20
|
|
|
|
|
+
|
|
|
|
|
+ 返回格式:
|
|
|
|
|
+ {
|
|
|
|
|
+ "success": True/False,
|
|
|
|
|
+ "result": {...} or None,
|
|
|
|
|
+ "message": "..."
|
|
|
|
|
+ }
|
|
|
|
|
+ """
|
|
|
|
|
+ # 1. 标注问题(仅一次)
|
|
|
|
|
+ annotation = await annotate_question(context.q_with_context)
|
|
|
|
|
+ context.question_annotation = annotation
|
|
|
|
|
+
|
|
|
|
|
+ # 2. 迭代优化
|
|
|
|
|
+ current_query = context.q
|
|
|
|
|
+
|
|
|
|
|
+ for round_num in range(1, max_rounds + 1):
|
|
|
|
|
+ print(f"\n{'='*60}")
|
|
|
|
|
+ print(f"第 {round_num} 轮:{'使用原始问题' if round_num == 1 else '使用修改后的query'}")
|
|
|
|
|
+ print(f"当前query: {current_query}")
|
|
|
|
|
+ print(f"{'='*60}")
|
|
|
|
|
+
|
|
|
|
|
+ # 获取推荐词并评估
|
|
|
|
|
+ evaluations = await get_suggestions_with_eval(current_query, annotation, context)
|
|
|
|
|
+
|
|
|
|
|
+ if evaluations:
|
|
|
|
|
+ # 检查是否找到合格query
|
|
|
|
|
+ qualified_queries = find_qualified_queries(evaluations, min_soft_score=0.7)
|
|
|
|
|
+ if qualified_queries:
|
|
|
|
|
+ return {
|
|
|
|
|
+ "success": True,
|
|
|
|
|
+ "results": qualified_queries,
|
|
|
|
|
+ "message": f"第{round_num}轮找到{len(qualified_queries)}个合格query"
|
|
|
|
|
+ }
|
|
|
|
|
+
|
|
|
|
|
+ # 如果是最后一轮,不再生成策略
|
|
|
|
|
+ if round_num == max_rounds:
|
|
|
|
|
+ break
|
|
|
|
|
+
|
|
|
|
|
+ # 生成修改策略
|
|
|
|
|
+ print(f"\n--- 生成修改策略 ---")
|
|
|
|
|
+ strategy = await generate_modification_strategy(annotation, context)
|
|
|
|
|
+
|
|
|
|
|
+ print(f"\n修改策略:")
|
|
|
|
|
+ print(f" 推理过程:{strategy.reasoning}")
|
|
|
|
|
+ print(f" 策略类型:{strategy.strategy_type}")
|
|
|
|
|
+ print(f" 基础query:{strategy.base_query} (来源: {strategy.base_query_source})")
|
|
|
|
|
+ print(f" 修改动作:{strategy.modification_action}")
|
|
|
|
|
+ print(f" 新query:{strategy.new_query}")
|
|
|
|
|
+
|
|
|
|
|
+ # 记录修改
|
|
|
|
|
+ context.operations_history.append({
|
|
|
|
|
+ "operation_type": "modify_query",
|
|
|
|
|
+ "timestamp": datetime.now().isoformat(),
|
|
|
|
|
+ "reasoning": strategy.reasoning,
|
|
|
|
|
+ "strategy_type": strategy.strategy_type,
|
|
|
|
|
+ "base_query": strategy.base_query,
|
|
|
|
|
+ "base_query_source": strategy.base_query_source,
|
|
|
|
|
+ "modification_action": strategy.modification_action,
|
|
|
|
|
+ "original_query": current_query,
|
|
|
|
|
+ "new_query": strategy.new_query,
|
|
|
|
|
+ })
|
|
|
|
|
+
|
|
|
|
|
+ # 更新当前query
|
|
|
|
|
+ current_query = strategy.new_query
|
|
|
|
|
+
|
|
|
|
|
+ # 所有轮次后仍未找到,从所有历史评估中降低标准查找
|
|
|
|
|
+ print(f"\n{'='*60}")
|
|
|
|
|
+ print(f"{max_rounds}轮后未找到最优query,降低标准(soft_score >= 0.5)")
|
|
|
|
|
+ print(f"{'='*60}")
|
|
|
|
|
+
|
|
|
|
|
+ # 收集所有历史轮次的评估结果
|
|
|
|
|
+ all_evaluations = []
|
|
|
|
|
+ for op in context.operations_history:
|
|
|
|
|
+ if op["operation_type"] == "get_query_suggestions" and op["evaluations"]:
|
|
|
|
|
+ all_evaluations.extend(op["evaluations"])
|
|
|
|
|
+
|
|
|
|
|
+ if not all_evaluations:
|
|
|
|
|
+ return {
|
|
|
|
|
+ "success": False,
|
|
|
|
|
+ "results": [],
|
|
|
|
|
+ "message": "所有轮次均未返回推荐词"
|
|
|
|
|
+ }
|
|
|
|
|
+
|
|
|
|
|
+ # 降级查找:soft_score >= 0.5
|
|
|
|
|
+ acceptable_queries = find_qualified_queries(all_evaluations, min_soft_score=0.5)
|
|
|
|
|
+ if acceptable_queries:
|
|
|
|
|
+ return {
|
|
|
|
|
+ "success": True,
|
|
|
|
|
+ "results": acceptable_queries,
|
|
|
|
|
+ "message": f"{max_rounds}轮后找到{len(acceptable_queries)}个可接受query(soft_score >= 0.5)"
|
|
|
|
|
+ }
|
|
|
|
|
+
|
|
|
|
|
+ # 完全失败:找出最接近的(essence=1, hard=1)
|
|
|
|
|
+ essence_hard_ok = [
|
|
|
|
|
+ e for e in all_evaluations
|
|
|
|
|
+ if e['essence_score'] == 1 and e['hard_score'] == 1
|
|
|
|
|
+ ]
|
|
|
|
|
+ if essence_hard_ok:
|
|
|
|
|
+ # 返回所有满足essence和hard的,按soft_score降序
|
|
|
|
|
+ closest_queries = sorted(essence_hard_ok, key=lambda x: x['soft_score'], reverse=True)
|
|
|
|
|
+ return {
|
|
|
|
|
+ "success": False,
|
|
|
|
|
+ "results": closest_queries,
|
|
|
|
|
+ "message": f"未找到合格query,但有{len(closest_queries)}个接近的推荐词(essence=1, hard=1)"
|
|
|
|
|
+ }
|
|
|
|
|
+
|
|
|
|
|
+ return {
|
|
|
|
|
+ "success": False,
|
|
|
|
|
+ "results": [],
|
|
|
|
|
+ "message": "未找到任何满足本质和硬性约束的推荐词"
|
|
|
|
|
+ }
|
|
|
|
|
+
|
|
|
|
|
+
|
|
|
|
|
+# ============================================================================
|
|
|
|
|
+# 输出格式化
|
|
|
|
|
+# ============================================================================
|
|
|
|
|
+
|
|
|
|
|
+def format_output(optimization_result: dict, context: RunContext) -> str:
|
|
|
|
|
+ """格式化输出结果"""
|
|
|
|
|
+ results = optimization_result.get("results", [])
|
|
|
|
|
+
|
|
|
|
|
+ if optimization_result["success"] and results:
|
|
|
|
|
+ output = f"原始问题:{context.q}\n"
|
|
|
|
|
+ output += f"状态:{optimization_result['message']}\n\n"
|
|
|
|
|
+ output += "合格的推荐query(按soft_score降序):\n"
|
|
|
|
|
+ for i, result in enumerate(results, 1):
|
|
|
|
|
+ output += f"\n{i}. {result['query']}\n"
|
|
|
|
|
+ output += f" - 本质匹配度:{result['essence_score']} (1=本质一致)\n"
|
|
|
|
|
+ output += f" - 硬性约束匹配度:{result['hard_score']} (1=所有约束满足)\n"
|
|
|
|
|
+ output += f" - 软性修饰完整度:{result['soft_score']:.2f} (0-1)\n"
|
|
|
|
|
+ output += f" - 评估理由:{result['reason']}\n"
|
|
|
|
|
+ return output.strip()
|
|
|
|
|
+ else:
|
|
|
|
|
+ output = f"原始问题:{context.q}\n"
|
|
|
|
|
+ output += f"结果:未找到合格推荐query\n"
|
|
|
|
|
+ output += f"原因:{optimization_result['message']}\n"
|
|
|
|
|
+
|
|
|
|
|
+ if results:
|
|
|
|
|
+ output += "\n最接近的推荐词(按soft_score降序):\n"
|
|
|
|
|
+ for i, result in enumerate(results[:3], 1): # 只显示前3个
|
|
|
|
|
+ output += f"\n{i}. {result['query']}\n"
|
|
|
|
|
+ output += f" - essence_score: {result['essence_score']}\n"
|
|
|
|
|
+ output += f" - hard_score: {result['hard_score']}\n"
|
|
|
|
|
+ output += f" - soft_score: {result['soft_score']:.2f}\n"
|
|
|
|
|
+ output += f" - reason: {result['reason']}\n"
|
|
|
|
|
+
|
|
|
|
|
+ output += "\n建议:尝试简化问题或调整需求描述"
|
|
|
|
|
+ return output.strip()
|
|
|
|
|
+
|
|
|
|
|
+
|
|
|
|
|
+# ============================================================================
|
|
|
|
|
+# 主函数
|
|
|
|
|
+# ============================================================================
|
|
|
|
|
+
|
|
|
|
|
+async def main(input_dir: str, max_rounds: int = 20):
|
|
|
|
|
+ current_time, log_url = set_trace()
|
|
|
|
|
+
|
|
|
|
|
+ # 从目录中读取固定文件名
|
|
|
|
|
+ input_context_file = os.path.join(input_dir, 'context.md')
|
|
|
|
|
+ input_q_file = os.path.join(input_dir, 'q.md')
|
|
|
|
|
+
|
|
|
|
|
+ q_context = read_file_as_string(input_context_file)
|
|
|
|
|
+ q = read_file_as_string(input_q_file)
|
|
|
|
|
+ q_with_context = f"""
|
|
|
|
|
+<需求上下文>
|
|
|
|
|
+{q_context}
|
|
|
|
|
+</需求上下文>
|
|
|
|
|
+<当前问题>
|
|
|
|
|
+{q}
|
|
|
|
|
+</当前问题>
|
|
|
|
|
+""".strip()
|
|
|
|
|
+
|
|
|
|
|
+ # 获取当前文件名作为版本
|
|
|
|
|
+ version = os.path.basename(__file__)
|
|
|
|
|
+ version_name = os.path.splitext(version)[0]
|
|
|
|
|
+
|
|
|
|
|
+ # 日志保存目录
|
|
|
|
|
+ log_dir = os.path.join(input_dir, "output", version_name, current_time)
|
|
|
|
|
+
|
|
|
|
|
+ run_context = RunContext(
|
|
|
|
|
+ version=version,
|
|
|
|
|
+ input_files={
|
|
|
|
|
+ "input_dir": input_dir,
|
|
|
|
|
+ "context_file": input_context_file,
|
|
|
|
|
+ "q_file": input_q_file,
|
|
|
|
|
+ },
|
|
|
|
|
+ q_with_context=q_with_context,
|
|
|
|
|
+ q_context=q_context,
|
|
|
|
|
+ q=q,
|
|
|
|
|
+ log_dir=log_dir,
|
|
|
|
|
+ log_url=log_url,
|
|
|
|
|
+ )
|
|
|
|
|
+
|
|
|
|
|
+ # 执行优化流程(代码控制)
|
|
|
|
|
+ optimization_result = await optimize_query(run_context, max_rounds=max_rounds)
|
|
|
|
|
+
|
|
|
|
|
+ # 格式化输出
|
|
|
|
|
+ final_output = format_output(optimization_result, run_context)
|
|
|
|
|
+ print(f"\n{'='*60}")
|
|
|
|
|
+ print("最终结果")
|
|
|
|
|
+ print(f"{'='*60}")
|
|
|
|
|
+ print(final_output)
|
|
|
|
|
+
|
|
|
|
|
+ # 保存结果
|
|
|
|
|
+ run_context.final_output = final_output
|
|
|
|
|
+
|
|
|
|
|
+ # 保存 RunContext 到 log_dir
|
|
|
|
|
+ os.makedirs(run_context.log_dir, exist_ok=True)
|
|
|
|
|
+ context_file_path = os.path.join(run_context.log_dir, "run_context.json")
|
|
|
|
|
+ with open(context_file_path, "w", encoding="utf-8") as f:
|
|
|
|
|
+ json.dump(run_context.model_dump(), f, ensure_ascii=False, indent=2)
|
|
|
|
|
+ print(f"\nRunContext saved to: {context_file_path}")
|
|
|
|
|
+
|
|
|
|
|
+
|
|
|
|
|
+if __name__ == "__main__":
|
|
|
|
|
+ parser = argparse.ArgumentParser(description="搜索query优化工具")
|
|
|
|
|
+ parser.add_argument(
|
|
|
|
|
+ "--input-dir",
|
|
|
|
|
+ type=str,
|
|
|
|
|
+ default="input/简单扣图",
|
|
|
|
|
+ help="输入目录路径,默认: input/简单扣图"
|
|
|
|
|
+ )
|
|
|
|
|
+ parser.add_argument(
|
|
|
|
|
+ "--max-rounds",
|
|
|
|
|
+ type=int,
|
|
|
|
|
+ default=20,
|
|
|
|
|
+ help="最大迭代轮数,默认: 20"
|
|
|
|
|
+ )
|
|
|
|
|
+ args = parser.parse_args()
|
|
|
|
|
+
|
|
|
|
|
+ asyncio.run(main(args.input_dir, max_rounds=args.max_rounds))
|