Back to Product

AI System Architecture Design

AI System Architecture Design: Video Search Desktop Client

1. Overview

Objective: Design a native macOS application for offline video content retrieval using OCR-based text recognition (supporting English/Chinese).
Key Requirements:

  • Zero network dependency
  • High-accuracy OCR for video frames
  • Efficient indexing and search
  • Native macOS integration
  • Security and scalability

2. Architecture Diagram

┌──────────────────────┐       ┌─────────────────┐       ┌─────────────────┐  
│      User Interface  │──────▶│   Core Engine   │──────▶│   Data Storage  │  
│ (SwiftUI 5.0)        │◀──────│ (Swift/C++)     │◀──────│ (SQLite 3.38 +  │  
└──────────────────────┘       └─────────────────┘       │ Core Data)      │  
                               ▲        ▲                └─────────────────┘  
                               │        │  
                 ┌─────────────┘        └──────────────┐  
                 ▼                                     ▼  
       ┌───────────────────┐                ┌─────────────────────┐  
       │  OCR Processing   │                │   Search Engine     │  
       │ (Vision 4.0)      │                │ (Lucene 9.5 +       │  
       └───────────────────┘                │  Custom Tokenizers) │  
                                            └─────────────────────┘

3. Technology Stack & Versions

Component Technology Version Rationale
Frontend SwiftUI 5.0 Native macOS UX
Backend Core Swift 5.7 Performance + Apple ecosystem
OCR Engine Apple Vision Framework 4.0 On-device CN/EN text recognition
Search Indexing Apache Lucene 9.5 Offline inverted indexing
Database SQLite + Core Data 3.38 Local storage optimization
Video Processing AVFoundation - Native frame extraction
Concurrency Grand Central Dispatch (GCD) - Parallel processing

4. Component Breakdown

4.1. User Interface (SwiftUI 5.0)
  • Video Library Manager: Drag-and-drop video ingestion (MP4, MOV, MKV).
  • Search Dashboard: Keyword input with real-time suggestions.
  • Result Viewer: Thumbnail grid with timestamped OCR snippets.
  • Playback Panel: Integrated AVPlayer for clip preview.
4.2. Core Engine (Swift/C++)
  • Frame Extractor:
    • Uses AVAssetImageGenerator to sample frames at 1 FPS (configurable).
    • Dynamic resolution scaling (4K → 1080p) to reduce OCR workload.
  • OCR Pipeline:
    • VNRecognizeTextRequest for CN/EN text detection.
    • Confidence threshold: 90% (adjustable via settings).
    • Post-processing: Noise removal using regex filters.
  • Indexing Service:
    • Lucene-based inverted index mapping: (word → videoID + timestamp)
    • Custom tokenizers for Chinese (Jieba 0.47) and English (Snowball).
4.3. Search Engine (Lucene 9.5)
  • Query Processing:
    • Tokenization + stemming (Porter2 for EN, Jieba for CN).
    • Fuzzy matching (Levenshtein distance ≤ 2).
  • Ranking:
    • TF-IDF scoring boosted by temporal proximity.
  • Caching: LRU cache for frequent queries (min. 200ms latency).
4.4. Data Storage (SQLite 3.38 + Core Data)
  • Schema:
    • VideoMeta: path, duration, checksum (SHA-256).
    • OCRIndex: word, videoID, timestamps, confidence.
  • Encryption: AES-256 for metadata at rest.

5. Workflow

  1. Ingestion:
    • User adds video → checksum validation → extract key metadata.
  2. Processing:
    • Frame extraction → OCR via Vision → text normalization → Lucene indexing.
  3. Search:
    • Query tokenization → index lookup → relevance ranking → results rendering.
  4. Playback:
    • Direct frame seek using AVPlayer.seek(to: toleranceBefore: toleranceAfter:).

6. Performance Optimization

  • Parallelism:
    • GCD queues for OCR (4 threads) and indexing (2 threads).
  • Resource Management:
    • Frame batch processing (max 100 frames/batch).
    • Index compression (Zstandard).
  • Latency Targets:
    • Indexing: ≤ 1.5× real-time (e.g., 10-min video in ≤15 mins).
    • Search: ≤ 300ms for 10k-indexed videos.

7. Security & Privacy

  • Data Isolation: Sandboxed storage with macOS App Sandbox.
  • Permissions: Explicit user consent for video access.
  • No Telemetry: Zero data exfiltration; all processing on-device.

8. Scalability

  • Modular Design:
    • Plug-in architecture for future OCR engines (e.g., Tesseract).
  • Index Sharding: Splits index by video date/size.
  • Resource Scaling: Auto-reduces frame rate on low RAM.

9. Deployment

  • Build: Xcode 14.3, macOS SDK 13.0+.
  • Distribution: Notarized .dmg via App Store/standalone.
  • Dependencies: Embedded Lucene; Jieba via Swift Package Manager.

10. Metrics & Monitoring

  • Instrumentation:
    • os_signpost for profiling OCR/search latency.
    • Memory/CPU usage logs (disabled by default).
  • User-Configurable:
    • Frame sampling rate (0.5–5 FPS).
    • Index purge scheduler.

Character Count: 3,812
This architecture ensures offline efficiency, leverages Apple-native frameworks for optimal macOS integration, and scales for large video libraries while maintaining strict privacy standards.