AniSearch — AI-Powered Semantic Anime Discovery

AniSearch is a next-generation anime discovery engine. Instead of relying on exact keyword matches, it uses Natural Language Processing (NLP) and vector embeddings to let users search for anime based on plot descriptions, hyper-specific tropes, or general "vibes."

"A cyberpunk detective fighting rogue AI in a dystopian city" "Psychological horror in a school setting" "Overpowered MC destroys everyone"

---

Preview

Live preview - Click Here

---

Features

• Semantic Search — Describe what you want in plain English; no need to know exact titles
• 20,000+ Anime — Full dataset scraped from the Jikan API (MyAnimeList)
• Poster Art & Scores — Results display cover images, MAL ratings, episode counts, and genre tags
• Instant Suggestions — One-click suggestion chips to explore popular queries
• Responsive UI — Premium dark-themed interface built with React + Tailwind CSS

---

Architecture

---

Getting Started

Prerequisites

• Python 3.9+ with pip
• Node.js 18+ with npm

1. Clone the repo

git clone <your-repo-url>
cd Anime

2. Set up the backend

Option A: Run locally

# Install Python dependencies
pip install fastapi uvicorn sentence-transformers torch

# Generate embeddings (first time only — takes 3-5 min)
python semantic_search.py

# Start the API server
uvicorn server:app --reload --port 8000

Option B: Deploy to Hugging Face Spaces

The backend/ folder is a self-contained Docker app ready for HF Spaces:

# Push the backend/ folder to your HF Space repo
# It will auto-build using the Dockerfile and expose port 7860

3. Set up the frontend

cd frontend
npm install
npm run dev

Open http://localhost:5173 in your browser.

Note: By default the frontend points to the Hugging Face Space backend (https://abhay557-animesearch.hf.space). To use a local backend, update BACKEND_URL in src/App.jsx to http://localhost:8000.

---

API Reference

GET /search

Search for anime by natural language description.

Parameter	Type	Description
q	string	User search query (required)

Response:

{
  "results": [
    {
      "title": "Cowboy Bebop",
      "genres": "Action, Adventure, Drama, Sci-Fi",
      "synopsis": "In the year 2071...",
      "episodes": 26,
      "mal_score": 8.75,
      "image": "https://cdn.myanimelist.net/images/anime/4/19644l.webp",
      "score": 0.62
    }
  ]
}

• score — Cosine similarity (0–1) between query and anime description
• mal_score — MyAnimeList community rating (0–10)
• image — Poster art URL from MAL

---

How It Works

1. Data Collection — scrape.js streams all anime from the Jikan v4 API into jikan_full_dump.json
2. Embedding Generation — Each anime's title + genres + synopsis is encoded into a 384-dimensional vector using the all-MiniLM-L6-v2 model
3. Search — The user's query is encoded into the same vector space and compared against all anime using cosine similarity
4. Ranking — Top 12 results above a 0.20 confidence threshold are returned

---

Tech Stack

Layer	Technology
Frontend	React 19, Vite 6, Tailwind CSS 3, Lucide React
Backend	FastAPI, Uvicorn
AI/ML	Sentence Transformers (all-MiniLM-L6-v2), PyTorch
Data	Jikan API v4 (MyAnimeList), custom Node.js scraper
Hosting	Hugging Face Spaces (backend)

---

License

This project is for educational and personal use. Anime data is sourced from MyAnimeList via the Jikan API.