Semantic search
Search by meaning, not keywords. Powered by Postgres and pgvector for fast vector and keyword search.
Open source meme search engine
Find any meme by what's in it.
Meme Search uses AI to index your memes by their content and text, making them instantly retrievable with semantic search. All processing runs locally on your machine.
Your memes deserve better than scrolling through folders. Index them once with local AI, then search by meaning — not just filenames.
Features
Everything you need to organize and retrieve your memes.
Auto-generate descriptions
AI models describe your memes automatically. Target specific memes or bulk-generate descriptions for your entire collection.
Tags and filters
Create, edit, and assign tags. Filter by tags, directory paths, and description embeddings for precise control.
Manual editing
Edit or add descriptions manually for better search results. No need to wait for auto-generation.
Directory paths
Organize memes across multiple subdirectories. Rescan to automatically detect and index new additions.
Drag-and-drop upload
Upload memes directly through the web interface. Supports JPG, PNG, WEBP with bulk upload up to 50 files.
Dark mode
Toggle between light and dark themes for comfortable viewing in any environment.
Bulk generation
Generate descriptions for multiple memes at once for faster indexing of large collections.
Docker ready
One command to start. docker compose up pulls and runs the app, database, and AI generator.
How it works
Three services, one command, all local.
Point at your memes
Mount your meme directories via Docker volumes. The app detects and registers subdirectories through the settings UI.
AI describes them
A local vision-language model reads each meme and generates a text description. Choose from 6 models depending on your hardware.
Search by meaning
Descriptions are embedded as vectors in Postgres with pgvector. Search returns memes ranked by semantic similarity to your query.
AI Models
Choose the right model for your hardware.
All models run locally. Smaller models work on constrained hardware; larger ones produce richer descriptions.
Microsoft's compact vision-language model. Best balance of speed and quality for most setups.
Larger Florence variant for richer descriptions when you have the RAM to spare.
Hugging Face's lightweight model. Fast inference on minimal hardware.
Mid-range SmolVLM variant. Better descriptions than 256M with moderate resource use.
Full-size vision-language model. Best description quality, needs ~5GB memory.
Quantized Moondream2. Cuts memory to ~1.5-2GB with minimal quality loss. Ideal for CPU-only machines.
Install
One command to start searching your memes.
Requires Docker and Docker Compose. The app runs on port 3000 by default, customizable via .env file.
git clone https://github.com/neonwatty/meme-search.git
cd meme-search
docker compose up
# Open http://localhost:3000Tech stack
Built with Rails, Python, Postgres, and pgvector.
Rails 8 + Ruby 3.4
The web application, search interface, and meme management UI. Serves the frontend and orchestrates the pipeline.
Python 3.12
Runs the image-to-text models. Receives memes via API, returns generated descriptions back to Rails.
PostgreSQL 17 + pgvector
Stores meme metadata, descriptions, tags, and vector embeddings. Powers both keyword and semantic search.