mcpField
C
chatbot-analytics

chatbot-analytics

github.com/curiositech数据与数据库257210 次安装2 天前更新
代码仓库

关于 chatbot-analytics

chatbot-analytics 是用于为 AI 辅导聊天实现分析功能的技能，强调避免存储消息内容等禁止的敏感数据，介绍基于元数据的指标、聚合仪表盘、告警和测试。

主要功能

  1. 01注重 HIPAA 合规的对话指标
  2. 02基于元数据的类别检测
  3. 03会话与 Token 使用量汇总
  4. 04危机、错误和放弃率激增告警
  5. 05检查禁止数据的测试

适用场景

  1. 01衡量聊天参与度与完成情况
  2. 02监控 Token 用量和响应延迟
  3. 03不读取消息内容而跟踪类别趋势

安装

复制下面的命令，在终端中运行。

$npx skills add https://github.com/curiositech/some_claude_skills --skill chatbot-analytics

安装后在 Claude Code 中输入/chatbot-analytics确认已加载。

附带文件.claude-plugin/plugin.json

详细介绍

AI 聊天机器人分析

此技能可帮助你为 AI 教练聊天功能实现分析，同时确保符合 HIPAA 要求。

要跟踪的核心指标

根据行业最佳实践，跟踪以下 13 项关键指标：

指标描述符合 HIPAA 要求？
总会话数聊天会话数量是
平均每次会话的消息数每次对话的消息数是
平均会话时长聊天花费的时间是
参与率使用聊天功能的用户占比是
完成率自然结束的会话是
放弃率提前结束的会话是
响应时间AI 响应延迟是
Token 使用量消耗的 Token 总数/平均数是
错误率响应失败率是
回退率“我不明白”类响应的比率是
话题类别用户讨论的内容仅元数据
情绪趋势情绪变化方向仅衍生数据
危机触发紧急情况检测仅元数据

符合 HIPAA 要求的分析

应跟踪的内容
// Conversation metadata (SAFE)
interface ConversationAnalytics {
  id: string;
  conversationId: string;
  userId: string;  // For aggregation, not individual tracking
  startedAt: Date;
  endedAt: Date | null;
  messageCount: number;
  userMessageCount: number;
  aiMessageCount: number;
  totalTokens: number;
  inputTokens: number;
  outputTokens: number;
  category: string;  // Derived from metadata flags
  outcome: 'completed' | 'abandoned' | 'error' | 'crisis_escalated';
  avgResponseTime: number;
  hadFallback: boolean;
}
不应跟踪的内容
// NEVER store these in analytics
interface PROHIBITED {
  messageContent: string;      // PHI
  userQuery: string;           // PHI
  aiResponse: string;          // PHI
  specificTopics: string[];    // Could reveal health info
  exactSentiment: 'sad';       // Could reveal mental state
}

实现模式

跟踪对话开始
// src/lib/ai/analytics.ts
export async function trackConversationStart(
  conversationId: string,
  userId: string
): Promise<void> {
  await db.insert(conversationAnalytics).values({
    id: generateId(),
    conversationId,
    userId,
    startedAt: new Date(),
    messageCount: 0,
    totalTokens: 0,
    category: 'unknown',
    outcome: 'in_progress'
  });
}
跟踪消息交互
export async function trackMessageExchange(
  conversationId: string,
  tokens: { input: number; output: number },
  responseTimeMs: number,
  flags: { hadFallback: boolean; hasCrisisIndicator: boolean }
): Promise<void> {
  await db
    .update(conversationAnalytics)
    .set({
      messageCount: sql`message_count + 1`,
      totalTokens: sql`total_tokens + ${tokens.input + tokens.output}`,
      inputTokens: sql`input_tokens + ${tokens.input}`,
      outputTokens: sql`output_tokens + ${tokens.output}`,
      avgResponseTime: sql`(avg_response_time * (message_count - 1) + ${responseTimeMs}) / message_count`,
      hadFallback: flags.hadFallback,
      ...(flags.hasCrisisIndicator && { outcome: 'crisis_escalated' })
    })
    .where(eq(conversationAnalytics.conversationId, conversationId));
}
跟踪对话结束
export async function trackConversationEnd(
  conversationId: string,
  outcome: 'completed' | 'abandoned' | 'error'
): Promise<void> {
  await db
    .update(conversationAnalytics)
    .set({
      endedAt: new Date(),
      outcome
    })
    .where(eq(conversationAnalytics.conversationId, conversationId));
}

类别检测（基于元数据）

检测对话类别，但不读取内容：

// Categories based on metadata flags from AI response
interface AIResponseMetadata {
  usedCopingStrategies: boolean;
  usedCrisisProtocol: boolean;
  usedCheckInSupport: boolean;
  usedGeneralChat: boolean;
  requestedClarification: boolean;
}

function deriveCategory(metadata: AIResponseMetadata): string {
  if (metadata.usedCrisisProtocol) return 'crisis_support';
  if (metadata.usedCopingStrategies) return 'coping_strategies';
  if (metadata.usedCheckInSupport) return 'checkin_support';
  if (metadata.requestedClarification) return 'clarification';
  return 'general_chat';
}

仪表板汇总

会话指标
// Get aggregated session stats (HIPAA safe - no individual data)
async function getSessionStats(days: number = 30) {
  const since = subDays(new Date(), days);

  return db
    .select({
      totalSessions: count(),
      avgMessages: avg(conversationAnalytics.messageCount),
      avgDuration: avg(
        sql`JULIANDAY(ended_at) - JULIANDAY(started_at)) * 24 * 60`
      ),
      completionRate: sql`
        CAST(SUM(CASE WHEN outcome = 'completed' THEN 1 ELSE 0 END) AS FLOAT) /
        CAST(COUNT(*) AS FLOAT)
      `,
      crisisEscalations: sql`
        SUM(CASE WHEN outcome = 'crisis_escalated' THEN 1 ELSE 0 END)
      `
    })
    .from(conversationAnalytics)
    .where(gte(conversationAnalytics.startedAt, since));
}
用于成本跟踪的 Token 使用量
async function getTokenUsage(days: number = 30) {
  const since = subDays(new Date(), days);

  const result = await db
    .select({
      totalTokens: sum(conversationAnalytics.totalTokens),
      inputTokens: sum(conversationAnalytics.inputTokens),
      outputTokens: sum(conversationAnalytics.outputTokens),
      avgTokensPerSession: avg(conversationAnalytics.totalTokens)
    })
    .from(conversationAnalytics)
    .where(gte(conversationAnalytics.startedAt, since));

  // Estimate cost (Claude pricing)
  const inputCost = (result.inputTokens / 1_000_000) * 3.00;  // $3/M input
  const outputCost = (result.outputTokens / 1_000_000) * 15.00; // $15/M output

  return {
    ...result,
    estimatedCost: inputCost + outputCost
  };
}
类别细分
async function getCategoryBreakdown(days: number = 30) {
  const since = subDays(new Date(), days);

  return db
    .select({
      category: conversationAnalytics.category,
      count: count(),
      percentage: sql`
        CAST(COUNT(*) AS FLOAT) * 100.0 /
        (SELECT COUNT(*) FROM conversation_analytics WHERE started_at >= ${since})
      `
    })
    .from(conversationAnalytics)
    .where(gte(conversationAnalytics.startedAt, since))
    .groupBy(conversationAnalytics.category)
    .orderBy(desc(count()));
}

警报配置

针对值得关注的模式设置警报：

interface AnalyticsAlert {
  type: 'crisis_spike' | 'error_spike' | 'abandonment_spike';
  threshold: number;
  windowHours: number;
  action: 'log' | 'email' | 'slack';
}

const alerts: AnalyticsAlert[] = [
  {
    type: 'crisis_spike',
    threshold: 5,  // 5+ crisis escalations
    windowHours: 24,
    action: 'email'
  },
  {
    type: 'error_spike',
    threshold: 10, // 10+ errors
    windowHours: 1,
    action: 'slack'
  },
  {
    type: 'abandonment_spike',
    threshold: 0.5, // 50%+ abandonment rate
    windowHours: 24,
    action: 'log'
  }
];

数据库架构

CREATE TABLE conversation_analytics (
  id TEXT PRIMARY KEY,
  conversation_id TEXT NOT NULL,
  user_id TEXT NOT NULL,
  started_at TEXT NOT NULL,
  ended_at TEXT,
  message_count INTEGER DEFAULT 0,
  user_message_count INTEGER DEFAULT 0,
  ai_message_count INTEGER DEFAULT 0,
  total_tokens INTEGER DEFAULT 0,
  input_tokens INTEGER DEFAULT 0,
  output_tokens INTEGER DEFAULT 0,
  category TEXT DEFAULT 'unknown',
  outcome TEXT DEFAULT 'in_progress',
  avg_response_time REAL DEFAULT 0,
  had_fallback INTEGER DEFAULT 0,

  FOREIGN KEY (conversation_id) REFERENCES conversations(id),
  FOREIGN KEY (user_id) REFERENCES users(id)
);

CREATE INDEX idx_conv_analytics_started ON conversation_analytics(started_at);
CREATE INDEX idx_conv_analytics_user ON conversation_analytics(user_id);
CREATE INDEX idx_conv_analytics_outcome ON conversation_analytics(outcome);

测试分析

describe('Conversation Analytics', () => {
  it('tracks session without PHI', async () => {
    const analytics = await trackConversationStart('conv-123', 'user-456');

    // Verify no PHI is stored
    expect(analytics).not.toHaveProperty('messageContent');
    expect(analytics).not.toHaveProperty('userQuery');

    // Verify metadata is stored
    expect(analytics.conversationId).toBe('conv-123');
    expect(analytics.messageCount).toBe(0);
  });

  it('calculates aggregates correctly', async () => {
    const stats = await getSessionStats(30);

    expect(stats.totalSessions).toBeGreaterThanOrEqual(0);
    expect(stats.completionRate).toBeBetween(0, 1);
  });
});

资源

评论

支持 Markdown