Search Engine System

A scalable search engine system supporting both image and text search capabilities using vector similarity.

Image Search

Text Search

System Architecture

System Architecture

Features

• Image Search Engine: Search for similar images using deep learning embeddings

  • Detailed Documentation
  • Vector similarity search using Qdrant
  • Support for multiple image formats
  • Real-time image processing and embedding generation
  • Based on ResNet/EfficientNet architecture for feature extraction

• Text Search Engine: Advanced text search with Elasticsearch

  • Detailed Documentation
  • Dual search capabilities:
    • Autocomplete (Search-as-you-type) using Edge NGram Tokenizer
    • Full-text search with fuzzy matching
  • Custom scoring based on business metrics
  • Multi-field search across item and shop names
  • Support for Vietnamese language

Technical Details

Image Search Pipeline

• Preprocessing:

  • Image resizing and normalization
  • Data augmentation for training
  • Support for JPEG, PNG, and WebP formats

• Feature Extraction:

  • Deep CNN architectures (ResNet/EfficientNet)
  • ONNX format for cross-platform compatibility
  • TensorRT optimization for GPU inference
  • Output: 512/1024-dimensional embedding vectors

• Vector Storage & Search:

  • Qdrant vector database for efficient similarity search
  • HNSW index for fast approximate nearest neighbor search
  • Configurable distance metrics (cosine/euclidean)

Text Search Pipeline

• Text Processing & Analysis:

  • Custom Elasticsearch analyzers:
    • Keyword analyzer with lowercase and ASCII folding
    • Edge NGram analyzer for autocomplete (min_gram: 2, max_gram: 5)
    • Standard analyzer for full-text search
  • Character filters and tokenization
  • Support for Vietnamese text

• Search Approaches:

  1. Autocomplete (Search-as-you-type):

     • Edge NGram tokenizer for prefix matching
     • Custom completion suggester
     • Optimized for instant suggestions
     • Minimum 2 characters for suggestions

  2. Full-Text Search:

     • Multi-match query across fields:
       • item_name
       • shop_name
     • Fuzzy matching with AUTO fuzziness
     • Custom scoring based on business metrics:
       • Sale rate (discount percentage)
       • Sales volume (>1000 sales bonus)
       • Item price normalization

• Search Optimization:

  • Custom scoring template using Elasticsearch scripts
  • Batch indexing for efficient data ingestion
  • Asynchronous search operations
  • Configurable result size
  • Error handling and logging

• Elasticsearch Features:

  • Custom index mappings
  • Multiple field types and analyzers
  • Function score queries
  • Script-based scoring
  • Bulk indexing operations

Technology Stack

Model Serving

• NVIDIA Triton Inference Server:
  • Triton Server Documentation
  • Model versioning and A/B testing
  • Dynamic batching
  • Concurrent model execution
  • GPU optimization with TensorRT
  • Model format conversion pipeline:
    • PyTorch → ONNX → TensorRT

Infrastructure

• Containerization:

  • Docker multi-stage builds
  • Optimized container images
  • Docker Compose for development

• Orchestration:

  • Kubernetes deployment
  • Helm Charts for package management
  • Horizontal Pod Autoscaling
  • Resource management and scaling

• Monitoring & Logging:

  • Prometheus metrics
  • Grafana dashboards
  • Distributed tracing
  • Performance monitoring

Getting Started

1. Clone the repository:

git clone https://github.com/vectornguyen76/search-engine-system.git

2. Start the services using Docker Compose:

docker-compose up -d

3. Access the services:

• Image Search UI: http://localhost:8501
• Text Search UI: http://localhost:8502
• Triton Server: http://localhost:8000

Development

CI/CD Pipeline

• Development Environment:

  • Code linting (Flake8)
  • Unit tests
  • Integration tests

• Staging Environment:

  • Performance testing
  • Load testing
  • Security scanning

• Production Environment:

  • Blue-green deployment
  • Automated rollback
  • Performance monitoring

Code Quality

• Flake8 for Python code linting
• Type hints and documentation
• Automated testing in CI/CD pipeline
• Code review process

Contributing

1. Fork the repository
2. Create your feature branch
3. Commit your changes
4. Push to the branch
5. Create a new Pull Request

License

This project is licensed under the MIT License - see the LICENSE file for details.