MOCKI – AI-Powered Mock Interview Platform

Tech Stack: React + FastAPI + MongoDB + OpenAI + Qdrant + Redis + Celery + WebRTC + Firebase Auth + Resend

System Architecture

graph TD
    User[User] --> Frontend[Frontend React]
    Frontend --> WebRTC[WebRTC Audio/Video]
    Frontend --> Websocket[Websocket Real-time Communication]
    Frontend <--> FirebaseAuth[Firebase Auth Authentication]
 
    Websocket <--> NodeBackend[NestJS Backend]
    NodeBackend <--> RedisPS[Redis Pub/Sub]
    NodeBackend <--> PythonBackend[Python Backend]
 
    NodeBackend --> STT[Speech to Text STT]
    NodeBackend --> TTS[Text to Speech TTS]
    NodeBackend --> TranslationAPI[Translation API]
 
    PythonBackend --> OpenAI[OpenAI API]
    PythonBackend --> Qdrant[Qdrant Vector Database]
    PythonBackend --> MongoDB[MongoDB]
    PythonBackend --> Redis[Redis Cache]
    PythonBackend --> CeleryWorkers[Celery Workers]
 
    CeleryWorkers --> Flower[Flower Monitoring]
    Flower --> Prometheus[Prometheus]
    Prometheus --> Grafana[Grafana Visualization]
 
    PythonBackend --> Gmail[Gmail API]
    PythonBackend --> Resend[Resend Email Service]
 
    subgraph Security System
    RateLimiter[Redis Rate Limiter]
    BloomFilter[Bloom Filter]
    Vault[HashiCorp Vault]
    end
 
    PythonBackend --> Security System

🌟 Core Features & System Modules

1. AI-Powered Mock Interview Platform (Core Product Capability)

sequenceDiagram
    participant User
    participant Frontend
    participant NestJS
    participant PythonBackend
    participant OpenAI
    participant CacheSystem
 
    User->>Frontend: Initiate Interview Request
    Frontend->>NestJS: WebRTC Audio/Video Connection
    Frontend->>NestJS: Websocket Connection
    NestJS->>PythonBackend: Establish Websocket Connection
 
    loop Interview Process
        User->>Frontend: Audio Input
        Frontend->>NestJS: Audio Stream
        NestJS->>NestJS: STT Conversion
        NestJS->>PythonBackend: Text Question
 
        PythonBackend->>CacheSystem: Check Question Cache
        alt Cache Hit
            CacheSystem-->>PythonBackend: Return Cached Result
        else Cache Miss
            PythonBackend->>OpenAI: LLM Inference Request
            OpenAI-->>PythonBackend: Return Answer
            PythonBackend->>CacheSystem: Update Cache
        end
 
        PythonBackend->>NestJS: Return AI Answer
        NestJS->>NestJS: TTS Conversion
        NestJS->>Frontend: Audio Answer
        Frontend->>User: Play AI Answer
    end
 
    User->>Frontend: End Interview
    Frontend->>NestJS: Close Connection
    NestJS->>PythonBackend: Save Interview Record

Independently developed complete pipeline: audio-video interview process based on WebRTC → STT → LLM inference → TTS audio return, supporting seamless switching between AI and human interviewers
Built real-time communication system based on Socket.IO + Redis Pub/Sub + Redis Adapter, supporting message synchronization and state broadcasting in multi-client, multi-instance deployments
Supports role switching (system, AI, human), question bank management, resume import, and position-specific Q&A, building structured and unstructured question combinations
System Architecture:
- NestJS Backend: Handles audio-video calls and conversation records, HTTP calls to translation and TTS APIs, streaming STT (Baidu, Google, iFlytek), maintains WebSocket connections with frontend and Python backend
- Python Backend: Handles AI requests, RAG system, email analysis system, user login management, rate limiting. Connects with NodeJS via WebSocket in AI interview system, and interacts with frontend via REST API for login and email analysis

2. Intelligent Email Analysis System (AI Information Extraction + Auto-closing)

graph TD
    Start[User Authorization] --> OAuth[Gmail OAuth2.0 Authentication]
    OAuth --> FetchEmails[Fetch Application Emails]
    FetchEmails --> Filter[Filter by Labels/Time]
 
    Filter --> ParallelProcessing[Parallel Async Processing]
    ParallelProcessing --> CeleryTasks[Celery Task Distribution]
 
    CeleryTasks --> AnalyzeEmail1[Analyze Email 1]
    CeleryTasks --> AnalyzeEmail2[Analyze Email 2]
    CeleryTasks --> AnalyzeEmailN[Analyze Email N...]
 
    AnalyzeEmail1 --> LLMProcessing[LLM Information Extraction]
    AnalyzeEmail2 --> LLMProcessing
    AnalyzeEmailN --> LLMProcessing
 
    LLMProcessing --> ExtractStatus[Extract Application Status]
 
    ExtractStatus --> CreateTimeline[Generate Timeline]
    CreateTimeline --> GenerateTasks[Generate Tasks]
    CreateTimeline --> JobRecommendation[Job Recommendations]
 
    GenerateTasks --> Notification[Send Notifications]
    JobRecommendation --> Notification

Supports user login via Gmail OAuth2.0, uses Gmail API to fetch application emails, filters historical emails by labels/time
Uses parallel async LLM to analyze email content, extracts application status (read, replied, interview invitation), position name, company info, and aggregates into Timeline
Integrates company position information to push job recommendations to users

3. Todo Management & Smart Reminder System

Automatically generates todos based on email analysis and user behavior, "TODO" "DONE", building complete application tracking path
Uses Resend to implement time-triggered deadline reminders, real-time notifications, and status change pushes, avoiding key information omissions
Supports status synchronization and cross-device access, user operations instantly feedback to database

4. RAG System & Contextual Q&A Optimization

graph TD
    subgraph Index Construction
    Resume[User Resume] --> TextExtractor[Text Extraction]
    JobDesc[Job Description] --> TextExtractor
    TextExtractor --> Chunker[Text Chunking]
 
    Chunker --> QAPairCreator[Query-Answer Pair Generation]
    QAPairCreator --> QueryEmbedder[Query Embedding Generation]
    QAPairCreator --> AnswerEmbedder[Answer Embedding Generation]
 
    QueryEmbedder --> VectorStore[Qdrant Vector Storage]
    AnswerEmbedder --> VectorStore
    QAPairCreator --> MetadataStore[MongoDB Metadata]
    end
 
    subgraph Retrieval Process
    Query[User Question] --> InputEmbedder[Query Embedding Generation]
 
    InputEmbedder --> QuerySimilarity[Query Similarity Search]
    InputEmbedder --> AnswerSimilarity[Answer Similarity Search]
 
    QuerySimilarity --> Qdrant[(Qdrant\nVector Database)]
    AnswerSimilarity --> Qdrant
 
    QuerySimilarity --> WeightedMerger[Weighted Result Merging]
    AnswerSimilarity --> WeightedMerger
    WeightedMerger --> DocumentIDs[Return Similar Document IDs]
 
    DocumentIDs --> MongoDB[(MongoDB\nMetadata Storage)]
    MongoDB --> RetrievedDocs[Retrieve Complete Documents]
    end
 
    subgraph Context Management
    RetrievedDocs --> ContextFormatter[Context Formatting]
    ContextFormatter --> TokenCounter[Token Counter]
 
    PreviousQA[Previous Q&A] --> SummaryBuffer[Summary Buffer]
    SummaryBuffer --> TokenThreshold[Token Threshold Control]
    TokenThreshold --> ContextTruncation[Context Truncation]
    end
 
    subgraph Multi-Agent Routing
    Query --> IntentClassifier[Intent Classifier]
    IntentClassifier --> MemoryRouter[Memory Router]
    MemoryRouter --> BQAgent[BQ Interview Agent]
    MemoryRouter --> CodeAgent[Coding Interview Agent]
    MemoryRouter --> KnowledgeAgent[Basic Knowledge Agent]
    MemoryRouter --> AnalysisAgent[Interview Analysis Agent]
    MemoryRouter --> QuestionAgent[Question Agent]
    end
 
    ContextFormatter --> LLMPrompt[LLM Prompt Construction]
    ContextTruncation --> LLMPrompt
    MemoryRouter --> LLMPrompt
    LLMPrompt --> OpenAI[OpenAI API]
    OpenAI --> Response[Generate Response]

Vector search engine based on LangChain + Qdrant, supporting Query-Answer vector similarity search
Efficient retrieval process: user questions are vectorized and simultaneously compared with pre-processed Query and Answer vectors for similarity calculation, with weighted merging of results to obtain the most relevant content
Optimized document processing: interview materials are pre-processed into Query-Answer pairs, with separate embedding calculations, enabling retrieval to consider both question similarity and answer relevance
Simplified index structure: direct vector similarity sorting for retrieval, avoiding complex index algorithm tuning while meeting performance requirements at current data scale
Multi-turn Q&A using ConversationSummaryBufferMemory for summary caching, combined with Token threshold controller for automatic context truncation, reducing overall Token cost by 45%
Intent classification using LangChain LLMChain, driving MemoryRouter for dynamic context switching between multiple Agents
Professional Agent system includes:
- BQ Interview Agent: handling behavioral questions and soft skills assessment
- Coding Interview Agent: evaluating programming ability and algorithm analysis
- Basic Knowledge Agent: testing professional domain knowledge
- Interview Analysis Agent: providing interview performance evaluation and feedback
- Question Agent: generating in-depth questions based on candidate background

5. Asynchronous Task System (Performance Guarantee)

graph LR
    subgraph Task Generation
    WebRequest[Web Request] --> TaskRouter
    ScheduledJob[Scheduled Job] --> TaskRouter
    EmailAnalysis[Email Analysis] --> TaskRouter
    end
 
    subgraph Task Routing
    TaskRouter[Task Router] --> PriorityQueue
    TaskRouter --> UserPriority{User Priority}
    UserPriority -->|VIP User| HighPriority[High Priority Queue]
    UserPriority -->|Regular User| NormalPriority[Normal Priority Queue]
    end
 
    subgraph Queue System
    PriorityQueue[Priority Queue] --> Redis[(Redis)]
    HighPriority --> Redis
    NormalPriority --> Redis
    end
 
    subgraph Worker Pool
    Redis --> Dispatcher[Task Dispatcher]
    Dispatcher --> CPUWorkers[CPU-intensive Workers\nPDF/Email Parsing]
    Dispatcher --> IOWorkers[IO-intensive Workers\nLLM Calls/Embedding Generation]
    end
 
    subgraph Monitoring System
    CPUWorkers --> Flower[Flower Monitoring]
    IOWorkers --> Flower
    Flower --> Prometheus[Prometheus]
    Prometheus --> Grafana[Grafana]
    Grafana --> Alerts[Alert Notifications]
    end

Multi-machine multi-queue task scheduling system based on Celery + Redis, with Worker pool independently deployed for CPU-intensive (PDF/email parsing) and IO-intensive (LLM calls, embedding generation) tasks
Priority task scheduling (e.g., VIP user request queue jumping), reducing urgent task average latency by 60%
Real-time monitoring with Flower, combined with Prometheus/Grafana for automatic alerts and visual tracking of task failure rates and queue backlog metrics

Flower Monitoring Tool Details

Flower is Celery's officially recommended task monitoring tool, suitable for real-time monitoring of Celery task execution, Worker status, task queue backlog, etc.. It can be understood as Celery's visual Dashboard + API server.

Flower Features:

👨‍🔧 Worker status (active, stopped, heartbeat lost)
📦 Queue length (backlog status)
✅ Task execution status (success, failure, retry, duration)
🧩 Task details (input parameters, return results, exception stack)
📈 Real-time task throughput, failure rate charts

Start Flower:

celery -A your_app flower --port=5555

Alert Implementation with Prometheus + Grafana:

Use celery-exporter to expose Celery data as Prometheus metrics
Configure Prometheus rules (e.g., task failures > threshold)
Set trigger conditions in Grafana (e.g., failure rate > 10% within 5 minutes) for alerts (email/Slack/Feishu/Webhook)

6. Multi-dimensional Rate Limiting & Security Protection

graph TD
    Request[API Request] --> SecurityLayer[Security Layer]
 
    subgraph Rate Limiting System
    SecurityLayer --> RateLimiter[Rate Limiter]
    RateLimiter --> UserLevel[User-level Rate Limit\n10/min]
    RateLimiter --> IPLevel[IP-level Rate Limit\n50/min]
    RateLimiter --> GlobalLevel[Global OpenAI Call Limit\n500/min]
    end
 
    subgraph Replay Attack Prevention
    SecurityLayer --> RequestFingerprint[Request Fingerprint Generation]
    RequestFingerprint --> BloomFilter[Bloom Filter]
    BloomFilter --> Cache[(Redis Cache)]
    BloomFilter --> CheckDuplicate{Duplicate?}
    CheckDuplicate -->|Yes| Reject[Reject Request]
    CheckDuplicate -->|No| Accept[Accept Request]
    end
 
    subgraph Data Security
    UserData[User Sensitive Data] --> ClientEncryption[Client-side AES-256 Encryption]
    ClientEncryption --> SecureTransmission[Encrypted Transmission]
    SecureTransmission --> ServerStorage[Server Stores Only Ciphertext]
    KeyManagement[Key Management] --> Vault[(HashiCorp Vault)]
    end

User-level (10/min), IP-level (50/min), and global OpenAI call-level (500/min) rate limiting system built with Redis sliding window algorithm, intercepting 98% of high-frequency abuse requests
Combined with Redis Bloom filter to cache processed request fingerprints, preventing duplicate submissions or replay attacks, with false positive rate < 0.1%

7. Multi-level Cache & High-frequency Interface Performance Optimization

graph TD
    Request[API Request] --> CacheRouter[Cache Router]
 
    subgraph Level 1 Cache
    CacheRouter --> RedisCache[Redis Cache]
    RedisCache --> L1Hit{Hit?}
    L1Hit -->|Yes| L1Return[Return Result]
    L1Hit -->|No| L2Cache
 
    RedisData1[High-frequency User Data] --> RedisCache
    RedisData2[Recent Interview Summaries] --> RedisCache
    RedisData3[Popular Question Embeddings] --> RedisCache
    end
 
    subgraph Level 2 Cache
    L2Cache[MongoDB Query Cache] --> L2Hit{Hit?}
    L2Hit -->|Yes| L2Return[Return Result]
    L2Hit -->|No| DatabaseQuery
    end
 
    subgraph Database Query
    DatabaseQuery[Execute MongoDB Query] --> StoreInCache
    StoreInCache[Store in Cache] --> L1Update[Update Redis Cache]
    StoreInCache --> L2Update[Update Query Cache]
    end
 
    subgraph Cache Strategy
    CachePolicy[Cache Policy] --> LRU[LRU Eviction Strategy]
    CachePolicy --> TTL[TTL Expiration Strategy]
    LRU --> EvictOldData[Evict Old Data]
    TTL --> AutoExpire[Auto Expire]
    end

High-frequency data (e.g., user's last 3 interview summaries) using Redis cache with JSON serialization storage, reducing query latency from MongoDB's 15ms to 1ms
Pre-generated and cached high-frequency interview question embeddings, reducing OpenAI API call volume by 35%
Lightweight MongoDB query cache layer with LRU + TTL strategy for automatic old data eviction, reducing high-frequency query (e.g., user/position information) latency from 15ms to 1ms, increasing overall QPS by 4x

8. User Authentication & Data Security

前端集成 Firebase Auth 实现身份认证，后端结合 JWT 管理权限与登录态
敏感数据（如用户简历）使用客户端 AES-256 加密传输，服务端仅存储密文，密钥托管在 HashiCorp Vault

Performance Metrics & Optimization Results

graph LR
    subgraph Cache Performance Improvement
    MongoDB[MongoDB Query\n15ms] --> |After Optimization| RedisCache[Redis Cache\n1ms]
    APITokens[OpenAI API\nCall Volume] --> |After Optimization| APITokensReduced[OpenAI API\nCall Volume Reduced 35%]
    QPS[System QPS] --> |After Optimization| QPSIncreased[System QPS\nIncreased 4x]
    end
 
    subgraph Task Processing Optimization
    TaskDelay[Urgent Task Delay] --> |After Optimization| TaskDelayReduced[Urgent Task Delay\nReduced 60%]
    TokenCost[Token Cost] --> |After Optimization| TokenCostReduced[Token Cost\nReduced 45%]
    end
 
    subgraph Security Performance
    AbuseRequests[Abuse Requests] --> |Block Rate| BlockRate[Block 98%]
    FalsePositive[Bloom Filter\nFalse Positive Rate] --> FalsePositiveRate[<0.1%]
    end
 
    subgraph Retrieval Accuracy
    SearchAccuracy[Top-3 Retrieval\nResult Accuracy] --> AccuracyRate[92%]
    end

Technical Highlights

Real-time Audio-Video System
- Low-latency interview system built with WebRTC + Socket.IO + Redis Pub/Sub, supporting seamless switching between AI and human interviewers
High-performance RAG Architecture
- Vector search engine based on LangChain + Qdrant, supporting Query-Answer vector similarity search
Multi-level Cache Design
- Multi-level architecture with Redis + MongoDB query cache, reducing key query latency from 15ms to 1ms, increasing system QPS by 4x
Asynchronous Task Optimization
- CPU/IO separation deployment strategy based on Celery, with priority scheduling, reducing urgent task latency by 60%
Efficient Token Management
- Reducing overall token cost by 45% through summary caching and context truncation
Comprehensive Security Protection
- Multi-dimensional rate limiting + Bloom filter + end-to-end encryption, blocking 98% of abuse requests with <0.1% false positive rate
Multi-Agent Routing System
- Context-aware dynamic routing based on intent classification, supporting switching between specialized domain agents
Intelligent Email Analysis
- Gmail API integration + parallel async LLM processing + semantic extraction, building an intelligent tracking system for the entire application lifecycle

MOCKI – AI 驱动的模拟面试平台

技术栈：React + FastAPI + MongoDB + OpenAI + Qdrant + Redis + Celery + WebRTC + Firebase Auth + Resend

系统架构

graph TD
    User[用户] --> Frontend[前端 React]
    Frontend --> WebRTC[WebRTC 音视频]
    Frontend --> Websocket[Websocket 实时通信]
    Frontend <--> FirebaseAuth[Firebase Auth 认证]
 
    Websocket <--> NodeBackend[NestJS 后端]
    NodeBackend <--> RedisPS[Redis Pub/Sub]
    NodeBackend <--> PythonBackend[Python 后端]
 
    NodeBackend --> STT[语音转文本 STT]
    NodeBackend --> TTS[文本转语音 TTS]
    NodeBackend --> TranslationAPI[翻译 API]
 
    PythonBackend --> OpenAI[OpenAI API]
    PythonBackend --> Qdrant[Qdrant 向量数据库]
    PythonBackend --> MongoDB[MongoDB]
    PythonBackend --> Redis[Redis 缓存]
    PythonBackend --> CeleryWorkers[Celery Workers]
 
    CeleryWorkers --> Flower[Flower 监控]
    Flower --> Prometheus[Prometheus]
    Prometheus --> Grafana[Grafana 可视化]
 
    PythonBackend --> Gmail[Gmail API]
    PythonBackend --> Resend[Resend 邮件服务]
 
    subgraph 安全系统
    RateLimiter[Redis 限流器]
    BloomFilter[布隆过滤器]
    Vault[HashiCorp Vault]
    end
 
    PythonBackend --> 安全系统

🌟 核心功能与系统模块

1. AI 驱动的模拟面试平台（核心产品能力）

sequenceDiagram
    participant 用户
    participant 前端
    participant NestJS
    participant Python后端
    participant OpenAI
    participant 缓存系统
 
    用户->>前端: 发起面试请求
    前端->>NestJS: WebRTC音视频连接
    前端->>NestJS: Websocket连接
    NestJS->>Python后端: 建立Websocket连接
 
    loop 面试过程
        用户->>前端: 音频输入
        前端->>NestJS: 音频流
        NestJS->>NestJS: STT转换
        NestJS->>Python后端: 文本问题
 
        Python后端->>缓存系统: 检查问题缓存
        alt 缓存命中
            缓存系统-->>Python后端: 返回缓存结果
        else 缓存未命中
            Python后端->>OpenAI: LLM推理请求
            OpenAI-->>Python后端: 返回回答
            Python后端->>缓存系统: 更新缓存
        end
 
        Python后端->>NestJS: 返回AI回答
        NestJS->>NestJS: TTS转换
        NestJS->>前端: 音频回答
        前端->>用户: 播放AI回答
    end
 
    用户->>前端: 结束面试
    前端->>NestJS: 关闭连接
    NestJS->>Python后端: 保存面试记录

独立开发完整链路：音视频面试流程基于 WebRTC → STT → LLM 推理 → TTS 返回音频，支持 AI 面试官与真人面试官自由切换
构建基于 Socket.IO + Redis Pub/Sub + Redis Adapter 的实时通信系统，支持多客户端、多实例部署下的消息同步与状态广播
支持角色切换（系统、AI、真人）、问题库管理、简历导入与职位定向问答，构建结构化与非结构化问题组合
系统架构：
- NestJS后端：负责连接音视频通话和对话记录，HTTP调用翻译API和TTS API，流式STT（百度、谷歌、讯飞），维持与前端的websocket连接，同时与Python后端保持websocket连接
- Python后端：处理AI请求，RAG系统，邮件分析系统，管理用户登录，实现限流。在AI面试系统中与nodejs通过websocket连接，在登录、邮件分析系统与前端通过REST API交互

2. 邮件智能分析系统（AI 信息抽取 + 自动闭环）

graph TD
    Start[用户授权] --> OAuth[Gmail OAuth2.0认证]
    OAuth --> FetchEmails[拉取投递邮件]
    FetchEmails --> Filter[按标签/时间过滤]
 
    Filter --> ParallelProcessing[并行异步处理]
    ParallelProcessing --> CeleryTasks[Celery任务分配]
 
    CeleryTasks --> AnalyzeEmail1[分析邮件1]
    CeleryTasks --> AnalyzeEmail2[分析邮件2]
    CeleryTasks --> AnalyzeEmailN[分析邮件N...]
 
    AnalyzeEmail1 --> LLMProcessing[LLM信息提取]
    AnalyzeEmail2 --> LLMProcessing
    AnalyzeEmailN --> LLMProcessing
 
    LLMProcessing --> ExtractStatus[提取申请状态]
 
    ExtractStatus --> CreateTimeline[生成时间线]
    CreateTimeline --> GenerateTasks[生成待办事项]
    CreateTimeline --> JobRecommendation[职位推荐]
 
    GenerateTasks --> Notification[发送通知]
    JobRecommendation --> Notification

支持用户通过 OAuth2.0 登录 Gmail，使用 Gmail API 拉取投递邮件，按标签/时间过滤历史邮件
使用并行异步 LLM 分析邮件内容，提取申请状态（如已读、已回复、面试邀请）、职位名称、公司信息等字段，并聚合形成 Timeline
整合公司职位信息，向用户推送职位列表推荐

3. 待办事项管理与智能提醒系统

自动根据邮件解析与用户行为生成待办事项，"TODO" "DONE"等，构建完整投递跟踪路径
使用 Resend 实现基于时间触发的临期提醒、实时通知与状态变更推送，避免关键信息遗漏
支持状态同步与跨设备访问，用户操作即时反馈至数据库

4. RAG 系统与上下文问答能力优化

graph TD
    subgraph Index Construction
    Resume[用户简历] --> TextExtractor[文本提取]
    JobDesc[职位描述] --> TextExtractor
    TextExtractor --> Chunker[文本分块]
 
    Chunker --> QAPairCreator[Query-Answer对生成]
    QAPairCreator --> QueryEmbedder[Query嵌入生成]
    QAPairCreator --> AnswerEmbedder[Answer嵌入生成]
 
    QueryEmbedder --> VectorStore[Qdrant向量存储]
    AnswerEmbedder --> VectorStore
    QAPairCreator --> MetadataStore[MongoDB元数据]
    end
 
    subgraph Retrieval Process
    Query[用户问题] --> InputEmbedder[查询嵌入生成]
 
    InputEmbedder --> QuerySimilarity[Query相似度检索]
    InputEmbedder --> AnswerSimilarity[Answer相似度检索]
 
    QuerySimilarity --> Qdrant[(Qdrant\n向量数据库)]
    AnswerSimilarity --> Qdrant
 
    QuerySimilarity --> WeightedMerger[加权结果合并]
    AnswerSimilarity --> WeightedMerger
    WeightedMerger --> DocumentIDs[返回相似文档ID]
 
    DocumentIDs --> MongoDB[(MongoDB\n元数据存储)]
    MongoDB --> RetrievedDocs[检索完整文档]
    end
 
    subgraph Context Management
    RetrievedDocs --> ContextFormatter[上下文格式化]
    ContextFormatter --> TokenCounter[Token计数器]
 
    PreviousQA[历史对话] --> SummaryBuffer[摘要缓存]
    SummaryBuffer --> TokenThreshold[Token阈值控制]
    TokenThreshold --> ContextTruncation[上下文截断]
    end
 
    subgraph Multi-Agent Routing
    Query --> IntentClassifier[意图分类器]
    IntentClassifier --> MemoryRouter[记忆路由器]
    MemoryRouter --> BQAgent[BQ面试Agent]
    MemoryRouter --> CodeAgent[代码面试Agent]
    MemoryRouter --> KnowledgeAgent[基础知识Agent]
    MemoryRouter --> AnalysisAgent[面试分析Agent]
    MemoryRouter --> QuestionAgent[提问Agent]
    end
 
    ContextFormatter --> LLMPrompt[LLM提示构建]
    ContextTruncation --> LLMPrompt
    MemoryRouter --> LLMPrompt
    LLMPrompt --> OpenAI[OpenAI API]
    OpenAI --> Response[生成响应]

Vector search engine based on LangChain + Qdrant, supporting Query-Answer vector similarity search
Efficient retrieval process: user questions are vectorized and simultaneously compared with pre-processed Query and Answer vectors for similarity calculation, with weighted merging of results to obtain the most relevant content
Optimized document processing: interview materials are pre-processed into Query-Answer pairs, with separate embedding calculations, enabling retrieval to consider both question similarity and answer relevance
Simplified index structure: direct vector similarity sorting for retrieval, avoiding complex index algorithm tuning while meeting performance requirements at current data scale
Multi-turn Q&A using ConversationSummaryBufferMemory for summary caching, combined with Token threshold controller for automatic context truncation, reducing overall Token cost by 45%
Intent classification using LangChain LLMChain, driving MemoryRouter for dynamic context switching between multiple Agents
Professional Agent system includes:
- BQ Interview Agent: handling behavioral questions and soft skills assessment
- Coding Interview Agent: evaluating programming ability and algorithm analysis
- Basic Knowledge Agent: testing professional domain knowledge
- Interview Analysis Agent: providing interview performance evaluation and feedback
- Question Agent: generating in-depth questions based on candidate background

5. 异步任务系统（性能保障能力）

graph LR
    subgraph Task Generation
    WebRequest[Web Request] --> TaskRouter
    ScheduledJob[Scheduled Job] --> TaskRouter
    EmailAnalysis[Email Analysis] --> TaskRouter
    end
 
    subgraph Task Routing
    TaskRouter[Task Router] --> PriorityQueue
    TaskRouter --> UserPriority{User Priority}
    UserPriority -->|VIP User| HighPriority[High Priority Queue]
    UserPriority -->|Regular User| NormalPriority[Normal Priority Queue]
    end
 
    subgraph Queue System
    PriorityQueue[Priority Queue] --> Redis[(Redis)]
    HighPriority --> Redis
    NormalPriority --> Redis
    end
 
    subgraph Worker Pool
    Redis --> Dispatcher[Task Dispatcher]
    Dispatcher --> CPUWorkers[CPU-intensive Workers\nPDF/Email Parsing]
    Dispatcher --> IOWorkers[IO-intensive Workers\nLLM Calls/Embedding Generation]
    end
 
    subgraph Monitoring System
    CPUWorkers --> Flower[Flower Monitoring]
    IOWorkers --> Flower
    Flower --> Prometheus[Prometheus]
    Prometheus --> Grafana[Grafana]
    Grafana --> Alerts[Alert Notifications]
    end

Multi-machine multi-queue task scheduling system based on Celery + Redis, with Worker pool independently deployed for CPU-intensive (PDF/email parsing) and IO-intensive (LLM calls, embedding generation) tasks
Priority task scheduling (e.g., VIP user request queue jumping), reducing urgent task average latency by 60%
Real-time monitoring with Flower, combined with Prometheus/Grafana for automatic alerts and visual tracking of task failure rates and queue backlog metrics

Flower Monitoring Tool Details

Flower Features:

👨‍🔧 Worker status (active, stopped, heartbeat lost)
📦 Queue length (backlog status)
✅ Task execution status (success, failure, retry, duration)
🧩 Task details (input parameters, return results, exception stack)
📈 Real-time task throughput, failure rate charts

Start Flower:

celery -A your_app flower --port=5555

Alert Implementation with Prometheus + Grafana:

Use celery-exporter to expose Celery data as Prometheus metrics
Configure Prometheus rules (e.g., task failures > threshold)
Set trigger conditions in Grafana (e.g., failure rate > 10% within 5 minutes) for alerts (email/Slack/Feishu/Webhook)

6. 多维限流与安全防护机制

graph TD
    Request[API Request] --> SecurityLayer[Security Layer]
 
    subgraph Rate Limiting System
    SecurityLayer --> RateLimiter[Rate Limiter]
    RateLimiter --> UserLevel[User-level Rate Limit\n10/min]
    RateLimiter --> IPLevel[IP-level Rate Limit\n50/min]
    RateLimiter --> GlobalLevel[Global OpenAI Call Limit\n500/min]
    end
 
    subgraph Replay Attack Prevention
    SecurityLayer --> RequestFingerprint[Request Fingerprint Generation]
    RequestFingerprint --> BloomFilter[Bloom Filter]
    BloomFilter --> Cache[(Redis Cache)]
    BloomFilter --> CheckDuplicate{Duplicate?}
    CheckDuplicate -->|Yes| Reject[Reject Request]
    CheckDuplicate -->|No| Accept[Accept Request]
    end
 
    subgraph Data Security
    UserData[User Sensitive Data] --> ClientEncryption[Client-side AES-256 Encryption]
    ClientEncryption --> SecureTransmission[Encrypted Transmission]
    SecureTransmission --> ServerStorage[Server Stores Only Ciphertext]
    KeyManagement[Key Management] --> Vault[(HashiCorp Vault)]
    end

User-level (10/min), IP-level (50/min), and global OpenAI call-level (500/min) rate limiting system built with Redis sliding window algorithm, intercepting 98% of high-frequency abuse requests
Combined with Redis Bloom filter to cache processed request fingerprints, preventing duplicate submissions or replay attacks, with false positive rate < 0.1%

7. 多级缓存与高频接口性能优化

graph TD
    Request[API Request] --> CacheRouter[Cache Router]
 
    subgraph Level 1 Cache
    CacheRouter --> RedisCache[Redis Cache]
    RedisCache --> L1Hit{Hit?}
    L1Hit -->|Yes| L1Return[Return Result]
    L1Hit -->|No| L2Cache
 
    RedisData1[High-frequency User Data] --> RedisCache
    RedisData2[Recent Interview Summaries] --> RedisCache
    RedisData3[Popular Question Embeddings] --> RedisCache
    end
 
    subgraph Level 2 Cache
    L2Cache[MongoDB Query Cache] --> L2Hit{Hit?}
    L2Hit -->|Yes| L2Return[Return Result]
    L2Hit -->|No| DatabaseQuery
    end
 
    subgraph Database Query
    DatabaseQuery[Execute MongoDB Query] --> StoreInCache
    StoreInCache[Store in Cache] --> L1Update[Update Redis Cache]
    StoreInCache --> L2Update[Update Query Cache]
    end
 
    subgraph Cache Strategy
    CachePolicy[Cache Policy] --> LRU[LRU Eviction Strategy]
    CachePolicy --> TTL[TTL Expiration Strategy]
    LRU --> EvictOldData[Evict Old Data]
    TTL --> AutoExpire[Auto Expire]
    end

High-frequency data (e.g., user's last 3 interview summaries) using Redis cache with JSON serialization storage, reducing query latency from MongoDB's 15ms to 1ms
Pre-generated and cached high-frequency interview question embeddings, reducing OpenAI API call volume by 35%
Lightweight MongoDB query cache layer with LRU + TTL strategy for automatic old data eviction, reducing high-frequency query (e.g., user/position information) latency from 15ms to 1ms, increasing overall QPS by 4x

8. 用户认证与数据安全

前端集成 Firebase Auth 实现身份认证，后端结合 JWT 管理权限与登录态
敏感数据（如用户简历）使用客户端 AES-256 加密传输，服务端仅存储密文，密钥托管在 HashiCorp Vault

性能指标与优化效果

graph LR
    subgraph Cache Performance Improvement
    MongoDB[MongoDB Query\n15ms] --> |After Optimization| RedisCache[Redis Cache\n1ms]
    APITokens[OpenAI API\nCall Volume] --> |After Optimization| APITokensReduced[OpenAI API\nCall Volume Reduced 35%]
    QPS[System QPS] --> |After Optimization| QPSIncreased[System QPS\nIncreased 4x]
    end
 
    subgraph Task Processing Optimization
    TaskDelay[Urgent Task Delay] --> |After Optimization| TaskDelayReduced[Urgent Task Delay\nReduced 60%]
    TokenCost[Token Cost] --> |After Optimization| TokenCostReduced[Token Cost\nReduced 45%]
    end
 
    subgraph Security Performance
    AbuseRequests[Abuse Requests] --> |Block Rate| BlockRate[Block 98%]
    FalsePositive[Bloom Filter\nFalse Positive Rate] --> FalsePositiveRate[<0.1%]
    end
 
    subgraph Retrieval Accuracy
    SearchAccuracy[Top-3 Retrieval\nResult Accuracy] --> AccuracyRate[92%]
    end

技术亮点总结

实时音视频交互系统
- WebRTC + Socket.IO + Redis Pub/Sub 构建的低延迟面试系统，支持AI与真人无缝切换
高性能RAG架构
- Vector search engine based on LangChain + Qdrant, supporting Query-Answer vector similarity search
多级缓存设计
- Redis + MongoDB查询缓存的多级架构，将关键查询延迟从15ms降至1ms，系统QPS提升4倍
异步任务优化
- 基于Celery的CPU/IO分离部署策略，结合优先级调度，紧急任务延迟降低60%
高效Token管理
- 通过摘要缓存与上下文截断技术，整体Token成本降低45%
全面的安全防护
- 多维限流 + 布隆过滤器 + 端到端加密，拦截98%滥用请求，误判率<0.1%
多Agent路由系统
- 基于意图分类的上下文动态路由，支持在不同专业领域Agent间切换
邮件智能分析
- Gmail API集成 + 并行异步LLM处理 + 语义提取，构建投递全生命周期的智能跟踪系统

MOCKI – AI-Powered Mock Interview Platform / MOCKI – AI 驱动的模拟面试平台

MOCKI – AI-Powered Mock Interview Platform

System Architecture

🌟 Core Features & System Modules

1. AI-Powered Mock Interview Platform (Core Product Capability)

2. Intelligent Email Analysis System (AI Information Extraction + Auto-closing)

3. Todo Management & Smart Reminder System

4. RAG System & Contextual Q&A Optimization

5. Asynchronous Task System (Performance Guarantee)

Flower Monitoring Tool Details

6. Multi-dimensional Rate Limiting & Security Protection

7. Multi-level Cache & High-frequency Interface Performance Optimization

8. User Authentication & Data Security

Performance Metrics & Optimization Results

Technical Highlights

MOCKI – AI 驱动的模拟面试平台

系统架构

🌟 核心功能与系统模块

1. AI 驱动的模拟面试平台（核心产品能力）

2. 邮件智能分析系统（AI 信息抽取 + 自动闭环）

3. 待办事项管理与智能提醒系统

4. RAG 系统与上下文问答能力优化

5. 异步任务系统（性能保障能力）

Flower Monitoring Tool Details

6. 多维限流与安全防护机制

7. 多级缓存与高频接口性能优化

8. 用户认证与数据安全

性能指标与优化效果

技术亮点总结