对话质量测试（Conversation Quality Testing）

# 对话质量评估最小示例
# 基于纯 Python，不依赖第三方库

from dataclasses import dataclass, field
from typing import List, Dict

@dataclass
class Turn:
    """单轮对话记录"""
    user: str                        # 用户输入
    agent: str                       # Agent 回复
    expected_intent: str = ""        # 期望识别的意图（用于意图准确率计算）
    needs_context: bool = False      # 这轮回复是否需要引用历史信息
    context_used: bool = False       # Agent 是否实际引用了历史信息

@dataclass
class EvalResult:
    """评估结果"""
    coherence: float = 0.0           # 连贯性得分 0-1
    intent_accuracy: float = 0.0     # 意图识别准确率 0-1
    context_retention: float = 0.0   # 上下文保持率 0-1
    overall: float = 0.0             # 加权总分

def evaluate_conversation(turns: List[Turn]) -> EvalResult:
    """
    对一段多轮对话执行质量评估。
    实际生产中，连贯性和意图识别通常交给 LLM-as-Judge 完成，
    此处用规则简化演示核心评估结构。
    """
    n = len(turns)
    if n == 0:
        return EvalResult()

    # 维度 1：连贯性（简化版：检测前后矛盾的关键词对）
    contradiction_pairs = [("没有库存", "现在有货"), ("不支持", "已经支持"), ("不知道", "答案是")]
    coherent = 0
    for i in range(1, n):
        prev, curr = turns[i-1].agent, turns[i].agent
        has_conflict = any(neg in prev and pos in curr for neg, pos in contradiction_pairs)
        if not has_conflict:
            coherent += 1
    coherence = coherent / (n - 1) if n > 1 else 1.0

    # 维度 2：意图识别准确率（对比期望意图和实际检测结果）
    def detect_intent(text: str) -> str:
        """基于关键词的简单意图分类"""
        if any(w in text for w in ["投诉", "问题", "坏了"]):
            return "complaint"
        if any(w in text for w in ["买", "购买", "下单"]):
            return "purchase"
        if any(w in text for w in ["什么", "怎么", "多少", "哪个"]):
            return "inquiry"
        return "other"

    correct = sum(1 for t in turns if t.expected_intent and detect_intent(t.user) == t.expected_intent)
    labeled = sum(1 for t in turns if t.expected_intent)
    intent_acc = correct / labeled if labeled > 0 else 1.0

    # 维度 3：上下文保持率（需要引用历史信息的轮次中，实际引用的比例）
    needs = [t for t in turns if t.needs_context]
    context_ret = sum(1 for t in needs if t.context_used) / len(needs) if needs else 1.0

    # 加权总分（权重可按业务调整）
    overall = coherence * 0.25 + intent_acc * 0.35 + context_ret * 0.40

    return EvalResult(
        coherence=round(coherence, 3),
        intent_accuracy=round(intent_acc, 3),
        context_retention=round(context_ret, 3),
        overall=round(overall, 3)
    )

# --- 构造测试数据并运行 ---

test_conversation = [
    Turn(user="iPhone 15 多少钱？", agent="iPhone 15 售价 5999 元。",
         expected_intent="inquiry"),
    Turn(user="它的配置呢？", agent="A17 芯片，6.1 英寸屏幕，支持 5G。",
         expected_intent="inquiry", needs_context=True, context_used=True),
    Turn(user="那个有黑色吗？", agent="iPhone 15 有黑色、蓝色、粉色三种配色。",
         expected_intent="inquiry", needs_context=True, context_used=True),
]

result = evaluate_conversation(test_conversation)
print(f"连贯性: {result.coherence}")
print(f"意图准确率: {result.intent_accuracy}")
print(f"上下文保持率: {result.context_retention}")
print(f"加权总分: {result.overall}")
# 输出:
# 连贯性: 1.0
# 意图准确率: 1.0
# 上下文保持率: 1.0
# 加权总分: 1.0