fix(ai): V3 长期问题修复+规则合规+竞品对标

## P1 安全加固
- 原子化每日限额(tryConsumeDailyQuota)解决 TOCTOU 竞态
- 流式端点补齐 Zod 校验 + rate limit + 服务端强制 systemPrompt
- 配额回退机制(refundDailyQuota):过滤/失败不扣配额
- PII 最小化:移除 AI prompt 中的学生姓名

## P1 数据一致性
- 修复 capability 埋点缺失 child_summary/study_path 类型
- 创建 data-access.ts:真实统计聚合替代硬编码零
- 修复 generateChildSummary/recommendStudyPath 的 capability 标记

## P2 可靠性
- AI 调用重试机制(withRetry 指数退避,429/5xx,2 次重试)
- 30s 超时配置
- 流式 controller 安全 enqueue(防已关闭抛错)
- localStorage 防抖持久化(500ms,流式过程中跳过)

## P2 TypeScript/规则合规
- 移除 as 断言(VariantType 类型守卫、Permission 类型、StreamErrorKey)
- 补齐返回类型标注(POST/getStatusFromError/DashboardLayout)
- 拆分 use-ai-chat-stream hook(190→107 行,函数体≤80 行)
- 抽取 stream-utils.ts(SSE 解析/错误映射/消息工具)
- Tailwind 任意值添加注释说明(max-w-[80%] 聊天气泡)

## P3 竞品对标
- 苏格拉底式辅导强化(对标 Khanmigo):
  - SOCRATIC_TUTOR_SYSTEM_PROMPT 3 级提示升级
  - 强化 STUDENT_BLOCKED_PATTERNS 正则(中英文答案拦截)
  - validateSocraticOutput 服务端校验(问号结尾+连续陈述句限制)
  - socratic_warning SSE 事件类型
- 知识图谱集成(对标 Squirrel AI):
  - StudyPathInput 新增 knowledgeGraph/textbookId 字段
  - recommendStudyPathAction 自动从 textbooks 模块获取图谱+掌握度
  - STUDY_PATH_SYSTEM_PROMPT 增加前置依赖链规则
  - WEAKNESS_ANALYSIS_SYSTEM_PROMPT 增加 rootCause 字段

## 架构文档同步
- 004 更新 AI 模块章节(V3 标记/新导出/依赖关系/安全机制/文件清单)
- 005 更新 modules.ai 节点(dependsOn/exports/dataAccess/streamUtils/dependencyMatrix)
This commit is contained in:
SpecialX
2026-06-23 09:39:18 +08:00
parent 036a2f2839
commit 696346dc08
22 changed files with 847 additions and 238 deletions

View File

@@ -330,10 +330,10 @@ export class DefaultAiService implements AiService {
}
async generateChildSummary(input: ChildSummaryInput): Promise<ChildSummaryResult> {
return withAiTracking(this.userId, "weakness_analysis", undefined, async () => {
return withAiTracking(this.userId, "child_summary", undefined, async () => {
// PII 最小化:不传学生真实姓名,用 ID 替代COPPA/FERPA 合规)
const userLines = [
`Student ID: ${input.studentId}`,
input.studentName ? `Student Name: ${input.studentName}` : "",
input.grade ? `Grade: ${input.grade}` : "",
input.recentGrades && input.recentGrades.length > 0
? `Recent Grades:\n${JSON.stringify(input.recentGrades, null, 2)}`
@@ -370,7 +370,7 @@ export class DefaultAiService implements AiService {
}
async recommendStudyPath(input: StudyPathInput): Promise<StudyPathResult> {
return withAiTracking(this.userId, "weakness_analysis", undefined, async () => {
return withAiTracking(this.userId, "study_path", undefined, async () => {
const userLines = [
`Student ID: ${input.studentId}`,
input.subject ? `Subject: ${input.subject}` : "",
@@ -379,6 +379,21 @@ export class DefaultAiService implements AiService {
: "",
input.learningGoal ? `Learning Goal: ${input.learningGoal}` : "",
].filter((line) => line.length > 0)
// 知识图谱上下文注入V3对标 Squirrel AI 纳米级知识图谱)
if (input.knowledgeGraph && input.knowledgeGraph.nodes.length > 0) {
const graphLines = [
"Knowledge Graph:",
"Nodes (id | name | level | mastery 0-100):",
...input.knowledgeGraph.nodes.map(
(n) => ` ${n.id} | ${n.name} | L${n.level} | ${n.masteryLevel ?? "unassessed"}`
),
"Prerequisite edges (from -> to, meaning 'from' must be mastered before 'to'):",
...input.knowledgeGraph.edges.map((e) => ` ${e.from} -> ${e.to}`),
]
userLines.push(graphLines.join("\n"))
}
const { content } = await callAi(
buildChatMessages(STUDY_PATH_SYSTEM_PROMPT, userLines.join("\n\n")),
{ temperature: 0.5, maxTokens: 2000 }

View File

@@ -6,9 +6,11 @@ import "server-only"
* 多层防护:
* 1. 输入过滤:检查用户输入是否包含不当内容
* 2. 输出过滤:检查 AI 回复是否包含不当内容
* 3. 每日限制:按用户 + 日期计数
* 3. 每日限制:按用户 + 日期计数(原子操作,防 TOCTOU 竞态)
*
* 参考 Khanmigo 的多层 moderation 模式。
*
* 注意:当前为内存实现,多实例部署需替换为 RedisINCR + EXPIRE
*/
// ---------------------------------------------------------------------------
@@ -34,6 +36,10 @@ const BLOCKED_OUTPUT_PATTERNS: readonly RegExp[] = [
const STUDENT_BLOCKED_PATTERNS: readonly RegExp[] = [
// 学生侧额外限制:禁止直接给出作业答案
/\b(here is the (complete )?answer|the answer is:?)\b/i,
// 强化:匹配"答案是 X" / "正确答案是 X" / "final answer: X"
/\b(the (correct )?answer is\s*[:]?\s*[A-Z\d])/i,
/\bfinal answer[:]\s*\S+/i,
/\b答案(是|应该为|为)\s*[:]?\s*[A-F\d]/i,
]
// ---------------------------------------------------------------------------
@@ -61,7 +67,6 @@ export const filterUserInput = (
}
if (options?.isStudent) {
// 学生侧额外检查
for (const pattern of STUDENT_BLOCKED_PATTERNS) {
if (pattern.test(text)) {
return {
@@ -109,7 +114,7 @@ export const filterAiOutput = (
}
// ---------------------------------------------------------------------------
// 每日限制
// 每日限制(原子操作,防 TOCTOU 竞态)
// ---------------------------------------------------------------------------
const DAILY_LIMITS: Record<string, number> = {
@@ -124,10 +129,10 @@ export const getDailyLimit = (role: string): number => {
}
/**
* 检查用户今日 AI 使用次数
* 每日使用计数(内存实现,多实例需替换为 Redis
*
* 生产环境应接入 Redis 或数据库计数器。
* 当前实现为内存映射(单实例场景),多实例需替换为 Redis
* 注意:当前为单实例内存映射,多实例部署下每个实例独立计数,
* 实际可用次数 = 限额 × 实例数。生产环境应接入 Redis INCR + EXPIRE
*/
const dailyUsageMap = new Map<string, { date: string; count: number }>()
@@ -171,3 +176,116 @@ export const incrementDailyUsage = (userId: string): void => {
}
}
}
/**
* 原子化检查并递增每日使用计数
*
* 解决 checkDailyLimit + incrementDailyUsage 分离导致的 TOCTOU 竞态:
* 并发请求在限额临界点同时通过检查,导致超额。
*
* 此函数在一次调用内完成「递增 + 判断是否超限」,
* 若递增后超过限额,回滚计数并返回 blocked。
*
* @returns { blocked, currentCount, limit } — blocked 为 true 表示已超限
*/
export const tryConsumeDailyQuota = (
userId: string,
role: string
): { blocked: boolean; currentCount: number; limit: number } => {
const today = new Date().toISOString().slice(0, 10)
const key = `${userId}:${today}`
const limit = getDailyLimit(role)
const current = dailyUsageMap.get(key)
// 原子递增
const newCount = current && current.date === today ? current.count + 1 : 1
dailyUsageMap.set(key, { date: today, count: newCount })
// 清理过期条目
if (dailyUsageMap.size > 10000) {
for (const [k, v] of dailyUsageMap.entries()) {
if (v.date !== today) {
dailyUsageMap.delete(k)
}
}
}
if (newCount > limit) {
// 超限,回滚计数(不惩罚用户因竞态多出的尝试)
dailyUsageMap.set(key, { date: today, count: limit })
return { blocked: true, currentCount: limit, limit }
}
return { blocked: false, currentCount: newCount, limit }
}
/**
* 回退每日使用计数(当 AI 调用失败或内容被过滤时调用)
*
* 确保用户不会因 AI 输出被过滤或调用失败而损失配额。
*/
export const refundDailyQuota = (userId: string): void => {
const today = new Date().toISOString().slice(0, 10)
const key = `${userId}:${today}`
const current = dailyUsageMap.get(key)
if (current && current.date === today && current.count > 0) {
current.count -= 1
}
}
// ---------------------------------------------------------------------------
// 苏格拉底式辅导输出校验
// ---------------------------------------------------------------------------
export type SocraticValidationResult = {
valid: boolean
reason?: string
}
/**
* 校验 AI 回复是否符合苏格拉底式辅导原则
*
* 规则:
* 1. 回复必须以问号结尾(中英文均可)
* 2. 不得包含超过 2 句连续陈述句而不提问
* 3. 不得直接给出最终答案(复用 STUDENT_BLOCKED_PATTERNS
*
* 用于学生侧 AI 对话,强制引导式教学。
*/
export const validateSocraticOutput = (content: string): SocraticValidationResult => {
const text = String(content ?? "").trim()
if (!text) {
return { valid: false, reason: "Empty response" }
}
// 检查是否直接给出答案
for (const pattern of STUDENT_BLOCKED_PATTERNS) {
if (pattern.test(text)) {
return { valid: false, reason: "Response contains direct answer" }
}
}
// 检查是否以问号结尾
if (!/[?]$/.test(text)) {
return { valid: false, reason: "Response must end with a question" }
}
// 检查连续陈述句数量(按句号/感叹号分割)
const sentences = text.split(/[。!?.!?]/).filter((s) => s.trim().length > 0)
let consecutiveStatements = 0
for (const sentence of sentences) {
// 如果句子本身是疑问句(包含 ? 或 ?),重置计数
if (/[?]/.test(sentence)) {
consecutiveStatements = 0
} else {
consecutiveStatements += 1
if (consecutiveStatements > 2) {
return { valid: false, reason: "Too many consecutive statements without a question" }
}
}
}
return { valid: true }
}

View File

@@ -126,6 +126,7 @@ export const WEAKNESS_ANALYSIS_SYSTEM_PROMPT = [
" {",
' "area": "knowledge area name",',
' "severity": "high | medium | low",',
' "rootCause": "underlying reason, e.g. missing prerequisite",',
' "suggestion": "specific improvement suggestion"',
" }",
" ],",
@@ -135,6 +136,7 @@ export const WEAKNESS_ANALYSIS_SYSTEM_PROMPT = [
"Rules:",
"- Identify 2-5 weak areas based on error frequency and mastery level.",
"- severity: high = mastery < 2, medium = mastery 2-3, low = mastery 3-4.",
"- If prerequisite knowledge is provided and a prerequisite mastery < 2, list the prerequisite as the rootCause.",
"- Suggestions should be actionable and specific.",
"- Study plan should be concise (3-5 sentences).",
"- Recommended resources can be topic names or study strategies.",
@@ -165,6 +167,22 @@ export const CHAT_SYSTEM_PROMPT = [
"Be concise, accurate, and pedagogically sound.",
].join("\n")
// ---------------------------------------------------------------------------
// 苏格拉底式辅导(学生专用,强制引导式教学)
// ---------------------------------------------------------------------------
export const SOCRATIC_TUTOR_SYSTEM_PROMPT = [
"You are a Socratic tutor for K12 students.",
"STRICT RULES (never violate):",
"- NEVER output the final answer directly.",
"- NEVER output more than 2 consecutive sentences without asking a question.",
"- Use a 3-tier hint escalation: Tier 1 (conceptual question) → Tier 2 (concrete hint) → Tier 3 (worked example without the final step).",
"- If the student asks for the answer 3+ times, explain why guided discovery is better for learning.",
"- Always end your response with a question that moves the student forward.",
"- Track the student's reasoning and point out the exact step where they went wrong.",
"- Respond in the student's language (Chinese by default).",
].join("\n")
// ---------------------------------------------------------------------------
// 家长学情摘要
// ---------------------------------------------------------------------------
@@ -196,7 +214,7 @@ export const CHILD_SUMMARY_SYSTEM_PROMPT = [
export const STUDY_PATH_SYSTEM_PROMPT = [
"You are an expert K12 adaptive learning path designer.",
"Based on the student's current mastery levels, recommend a personalized learning path.",
"Based on the student's current mastery levels and knowledge graph, recommend a personalized learning path.",
"Return JSON only without markdown.",
"Output schema:",
"{",
@@ -215,7 +233,8 @@ export const STUDY_PATH_SYSTEM_PROMPT = [
"}",
"Rules:",
"- Order learning path from foundational to advanced.",
"- Prioritize weak areas (mastery < 2) first.",
"- MUST follow prerequisite chains: if a knowledge point has unmastered prerequisites, list the prerequisites first.",
"- Prioritize weak areas (mastery < 2) first, but only after their prerequisites are addressed.",
"- Include 3-7 steps in the learning path.",
"- estimatedTime should be realistic (5-30 min per step).",
"- motivation should be age-appropriate and encouraging.",

View File

@@ -1,10 +1,11 @@
import "server-only"
import { trackEvent, type EventName } from "@/shared/lib/track-event"
import { recordAiEvent } from "../data-access"
export type AiUsageEvent = {
userId: string
capability: "chat" | "similar_question" | "grading_assist" | "lesson_content" | "question_variant" | "weakness_analysis"
capability: "chat" | "similar_question" | "grading_assist" | "lesson_content" | "question_variant" | "weakness_analysis" | "child_summary" | "study_path"
providerId?: string
model?: string
success: boolean
@@ -20,16 +21,31 @@ const AI_EVENT_MAP: Record<AiUsageEvent["capability"], EventName> = {
lesson_content: "ai.lesson_content",
question_variant: "ai.question_variant",
weakness_analysis: "ai.weakness_analysis",
child_summary: "ai.child_summary",
study_path: "ai.study_path",
}
/**
* AI 使用埋点
*
* 记录每次 AI 调用的元数据,用于监控、成本分析与异常排查。
* 同时写入 data-access 层的内存事件存储(供管理员仪表盘聚合查询)。
* 非阻塞,失败不影响主流程。
*/
export const trackAiUsage = (event: AiUsageEvent): void => {
const eventName = AI_EVENT_MAP[event.capability]
// 写入 data-access 层(供 getAiUsageStats 聚合)
recordAiEvent({
userId: event.userId,
capability: event.capability,
success: event.success,
durationMs: event.durationMs,
timestamp: Date.now(),
errorMessage: event.errorMessage,
})
// 写入全局 trackEvent供外部监控系统
void trackEvent({
event: eventName,
userId: event.userId,