🧠 Mangakensaku

Mangakensaku is a simple yet powerful FastAPI-based application that takes a text query and retrieves the most relevant manga panels. It was built primarily to understand how FAISS works for similarity search over image embeddings.

🚀 Features

• 🔍 Search manga panels using natural language queries
• 🖼️ Retrieve semantically similar images using CLIP embeddings
• ⚡ Fast approximate nearest neighbor search powered by FAISS
• ✅ Normalized embeddings + IndexFlatIP for best performance

🛠️ Tech Stack

• FastAPI for API framework
• FAISS for efficient similarity search
• CLIP for text & image embeddings
• Python for the backend logic

🧪 How It Works

1. A query is embedded using the CLIP model.
2. The embedding is normalized to unit length.
3. FAISS retrieves top-k most relevant images using IndexFlatIP.
4. Relevant manga panel URLs or paths are returned.

Note: FAISS IndexFlatIP works best with normalized embeddings because it uses inner product (which becomes cosine similarity when vectors are normalized).

🧰 Setup

git clone https://github.com/yourusername/mangakensaku.git
cd mangakensaku
pip install -r requirements.txt
uvicorn app.main:app --host 0.0.0.0 --port 10000

Image Credit

https://panelsdesu.com/