celt313/agentic-rag-gamequest

GameQuest

Overview

GameQuest is an intelligent game recommendation system that leverages multi-modal vector search and Agentic RAG technologies to help users discover their next favorite game. The system integrates multiple AI models including SentenceTransformers for semantic search, DINOv2 for visual similarity, and Qwen3 LLM for intelligent reasoning, providing both traditional filtering and advanced AI-powered recommendations with detailed explanations.

Features

• —Advanced Filtering: Platform, genre, score, year, and scored-only filters
• —Text Search: Find games by description, title, or metadata
• —Semantic Search: Vector-based similarity search using SentenceTransformers embeddings
• —AI Agentic RAG: Intelligent recommendations with reasoning using Qwen3 LLM
• —Visual Search: Find similar games by cover art or screenshots using DINOv2
• —Interactive UI: Clickable game cards with detailed modals

Architecture

Frontend (Flask/Gradio)
↓
Core Backend
├── Models (SentenceTransformers, DINOv2, Qwen3)
├── Retrieval (Vector Search, Reranking)
├── Database (PostgreSQL + ChromaDB)
└── AI Agent (Agentic RAG)

Project Structure

gamequest/
├── flask/              # Flask web application
│   ├── app.py         # Main Flask app
│   ├── static/        # CSS, JS, images
│   └── templates/     # HTML templates
├── gradio/             # Gradio interface
│   ├── app.py         # Main Gradio app
│   ├── components.py  # UI components
│   └── search_handlers.py # Search logic
├── core/               # Shared backend logic
│   ├── models/        # Model loading and inference
│   ├── retrieval/     # Search services and RAG
│   ├── utils/         # Database utilities
│   └── database/      # Database migration scripts
├── data/               # Database files and embeddings
├── notebooks/          # Jupyter notebooks for analysis
├── Dockerfile          # Container configuration
├── requirements.txt    # Python dependencies
└── README.md

Live Demo

🚀 Try GameQuest live: Hugging Face Spaces Demo

Search Capabilities

Text Search

• —Search by game description, title, or metadata
• —Filter by platform, genre, score, year
• —Supports both exact and fuzzy matching

Semantic Search

• —Vector-based similarity search
• —Uses SentenceTransformers embeddings
• —Finds conceptually similar games

AI Agentic RAG

• —Intelligent recommendations with reasoning
• —Uses Qwen3 LLM for natural language understanding
• —Provides detailed explanations for recommendations

Visual Search

• —Upload cover art or screenshots
• —Find visually similar games using DINOv2
• —Supports both cover and screenshot similarity

Data Sources

• —188,000+ games from MobyGames database
• —600,000+ critic reviews and ratings
• —Vector embeddings for semantic search
• —Image embeddings for visual similarity
• —Structured metadata (platforms, genres, scores, years)

Tech Stack

• —Frontend: HTML5, CSS3, JavaScript (Flask) / Gradio
• —Backend: Python 3.11+, Flask/Gradio
• —AI/ML: Hugging Face Transformers, SentenceTransformers, DINOv2
• —Database: PostgreSQL (Neon), ChromaDB
• —LLM: Qwen3 (Ollama local / HF Transformers cloud)
• —Deployment: Docker, Hugging Face Spaces

Configuration

• —SentenceTransformers: all-MiniLM-L6-v2 for text embeddings
• —DINOv2: dinov2-base for image embeddings
• —Reranker: BAAI/bge-reranker-base for result ranking
• —LLM: Qwen3-4B (cloud) / qwen3:0.6b (local)

Use Cases

• —Game Discovery: Find games by description or visual similarity
• —AI Recommendations: Get intelligent suggestions with reasoning
• —Research: Explore game database with advanced filtering
• —Development: Build on top of the vector search infrastructure

License

MIT License - see LICENSE file for details.