meme-search/meme-search

What it solves

Meme Search provides a way to organize and retrieve a personal collection of memes using AI. It eliminates the need to manually name files or remember where a specific image is stored by indexing memes based on their visual content and text.

How it works

The system uses image-to-text models to automatically generate descriptions of memes. These descriptions, along with manual edits and tags, are converted into vector embeddings and stored in a PostgreSQL database with the pgvector extension. This allows users to perform both keyword-based and semantic (vector) searches to find images. Users can choose between local inference using bundled Python services (with models like Florence-2 or Moondream2) or an external OpenAI-compatible vision API.

Who it’s for

Individuals who have large libraries of memes and want a self-hosted, private search engine to manage and retrieve them quickly.

Highlights

• Flexible AI Models: Supports various local vision-language models (Florence-2, SmolVLM, Moondream2) with different parameter sizes to fit different hardware.
• Hybrid Search: Combines traditional keyword search with fast vector search for more accurate results.
• Self-Hosted & Private: Designed to run via Docker, keeping image processing and embeddings local by default.
• Extensible Access: Includes a read-only Search API v1, a Python CLI, and an experimental browser extension for accessing the meme library outside the web UI.
• Bulk Processing: Supports bulk description generation and drag-and-drop uploads for fast indexing.