🔍 Semantic Image Search Engine

A high-performance, multimodal search engine powered by OpenAI CLIP and Qdrant Vector Database.

📌 Overview

This project implements a semantic search engine that allows users to find images using natural language descriptions (e.g., "a peaceful mountain lake at sunset") or by uploading a reference image.

Unlike traditional keyword search, this system understands the content and context of images using high-dimensional vector embeddings generated by the CLIP (Contrastive Language-Image Pre-training) model.

🏗️ System Architecture

The application follows a modern microservices pattern:

1. Frontend (Streamlit): Lightweight UI for user interaction and result visualization.
2. Backend (FastAPI): Handles model inference (CPU/GPU) and coordinates with the database.
3. Database (Qdrant): Stores 512-dimensional vector embeddings and metadata.

---

🚀 Getting Started

1️⃣ Prerequisites

Before running the project, ensure you have the following installed:

• Docker Desktop (Required to run Qdrant and the Microservices).
• Python 3.13+ (For running local setup scripts).
• Unsplash Lite Dataset:
  • Download the Unsplash Lite Dataset (25k images) freely available online.
  • Extract it to a folder on your machine (e.g., Downloads/unsplash-lite).

2️⃣ Configuration

You must update the file paths to point to your downloaded dataset.

1. Update docker-compose.yml: Change the volume path under services: -> backend: and services: -> frontend: to match your local image folder.

   volumes:
      - C:/Path/To/Your/unsplash-lite:/data/images:ro

2. Update qdrant_setup.py: Update the IMAGE_FOLDER variable at the top of the file:

   IMAGE_FOLDER = r"C:\Path\To\Your\unsplash-lite"

3️⃣ Installation & Run

1. Clone the repository:

   git clone https://github.com/yourusername/semantic-image-search.git
   cd semantic-image-search

2. Start the Services: Run the following command to build and start the Docker containers:

   docker-compose up --build

   Wait until you see the logs Uvicorn running on http://0.0.0.0:8000 and Streamlit ... http://0.0.0.0:8501.

4️⃣ Data Ingestion (First Run Only)

Once the Docker containers are running (specifically the Qdrant database on port 6333), you need to populate the database with your images.

Open a new terminal window and run:

# Make sure you have the dependencies installed locally first
pip install -r requirements.txt

# Run the ingestion script
python qdrant_setup.py

5️⃣ Access the Application

• Frontend (UI): http://localhost:8501 - Open this in your browser to use the search engine.
• Backend (API Docs): http://localhost:8000/docs - Swagger UI for testing API endpoints.
• Qdrant Dashboard: http://localhost:6333/dashboard - View your vector collection status.

---

💾 Data Persistence & Production Notes

This project uses Docker Bind Mounts (./qdrant_data:/qdrant/storage) for simplicity and portability. This ensures:

1. Zero-Config: You can clone the repo and run it without setting up external database clouds.
2. Easy Backup: All your vector data is visible in the qdrant_data folder on your host machine.

Note for Production: In a large-scale enterprise environment, this would typically be replaced by Docker Named Volumes (for better I/O performance) or a Managed Vector Database (like Qdrant Cloud or AWS RDS) to handle backups and high availability.

---

🛠️ Tech Stack

Component	Technology	Description
Model	OpenAI CLIP	Multimodal (Text-to-Image) Embeddings
Backend	FastAPI	High-performance Async API
Frontend	Streamlit	Data Science & ML Web Interface
Database	Qdrant	Vector Search Engine (Running in Docker)
DevOps	Docker Compose	Container Orchestration

---

📈 Performance

• Dimension: 512
• Search Latency: < 50ms (Approx, depends on hardware)
• Model: ViT-B/16 (Running on Python 3.13)