Semantic Image Search with DocArray + Transformers

Semantic Pet Vision is a lightweight semantic image search engine that retrieves similar pet images (cats 🐱 and dogs 🐶) based on their visual meaning, not just filenames or keywords.

By leveraging Hugging Face vision encoders and DocArray, the system transforms images into embeddings and performs similarity search. A simple query like: "cute ginger kitten"

returns the most semantically relevant image from the dataset.

1. Indexing – Images are encoded into embeddings using CLIP.
2. Storage – Embeddings are stored in doc.tags using DocArray.
3. Search – Natural language queries are encoded & compared with cosine similarity.
4. Result – The closest image is returned as top_match.jpg.

• Semantic Search → go beyond keywords, search by meaning
• Pet Focused → trained on cats & dogs for fun experimentation
• DocArray Powered → efficient storage + similarity search
• MIT Licensed → free to use, modify, and expand

Run the main script to:

1. Create documents from sample images
2. Generate embeddings
3. Perform a query search

python main.py

Example query:

Query: cute ginger kitten
Top match image saved to top_match.jpg

Though fun-sized, this project is a microcosm of real-world systems.

• 🔍 Content-based image retrieval (e.g., “find me similar products”)
• 🐕 Pet adoption search engines (search by photo)
• 📷 Duplicate detection in large photo collections

This project demonstrates semantic retrieval in computer vision, bridging representation learning and information retrieval. It’s a practical, compact example of applying deep embeddings for multimodal search tasks.

• Expand dataset beyond cats & dogs
• Integrate a web-based UI for interactive search
• Experiment with multi-modal queries (text + image)

This project is licensed under the MIT License - free for academic, and personal use.