Semantic search for your unlabeled meme hoard

What it does

Meme Search is a self-hosted web app that ingests your image folders—JPG, PNG, WEBP—and builds a searchable index from their actual visual content. It runs a Rails frontend and a Python microservice that feeds local vision-language models like Florence-2 or Moondream2 to auto-generate descriptions, then stores vector embeddings in Postgres with the pgvector extension. You query with plain text and get back semantically relevant memes, or fall back to keyword search and tag filters if you prefer precision over vibes.

The interesting bit

The project treats meme management as a first-class infrastructure problem: it bundles model selection, INT8 quantization for low-RAM machines, bulk background jobs via Solid Queue, and even drag-and-drop uploads into a single Docker Compose stack. That makes it less a toy and more a homelab appliance—an oddly practical use case for vector search in a category usually dominated by cloud photo apps.

Key highlights

• Runs fully offline by default; optional OpenAI-compatible vision API if you want cloud descriptions while keeping embeddings local.
• Swappable vision models from 256M to 2B parameters, including an INT8-quantized variant for CPU-only or memory-constrained hosts.
• Vector and keyword search with filtering by tags, directory paths, and description embeddings.
• Bulk description generation, directory rescanning, and drag-and-drop web uploads (up to 50 files at once).
• Dark mode, because even meme archivists deserve eye strain relief.

Verdict

Worth spinning up if you maintain a large local meme collection and want semantic retrieval without sending your images to a cloud service. Skip it if you already store everything in a managed photo platform with built-in AI search.