Image Search Engine (CLIP)

A lightweight, reproducible semantic image search pipeline using a SOTA CLIP embeddings.
It downloads images from URLs, computes their vector representations, and lets you search by text, image, or precomputed embeddings.

Examples

Text → Image search

Example query: "Netanyahu is committing a genocide"

---

Image → Image search

Example query: aerial destruction photo (Gaza, 2024)

---

Features

• Parallel download with retries + error logging (parallel_download, download_with_retry)
• SOTA CLIP embeddings (laion/CLIP-ViT-H-14-laion2B-s32B-b79K )
• Reusable parquet files for fast reloads
• Text → Image and Image → Image semantic search
• Simple visualization utilities (show_image, search_engine)

---

Datasets (examples)

1. 10 septembre — ~41,942 image URLs
2. Reconnaissance de la Palestine — ~5,055 image URLs

Only URLs are stored in the repo — images are downloaded locally.

---

Project Structure

image_search_engine/ │ ├─ assets/ │ ├─ image_search_text.png │ ├─ image_search_image.png ├─ img_urls/ # {subject}.csv files ├─ img_data/ # downloaded images (ignored) ├─ img_embeddings/ # parquet embeddings ├─ .env ├─ .gitignore │ ├─ helper_functions.py ├─ main.ipynb ├─ README.md └─ requirements.txt

data/img_data/ is inside .gitignore

Quick Start

# install
pip install -r requirements.txt

# run main notebook
1 main.ipynb

Remark

The transformer-based backbone (CLIP, DINO, VIT...) follows this pipeline: -image is divided into patches → linear projection (independent embeddings) → positional encoding added → attention mechanism (patches "communicate") → final aggregation (mean or [CLS]) → global vector.