meme-search

Why we included this project

Searching a large personal image collection usually means scrolling through folders and hoping the filenames mean something. This project turns that pile into a searchable library by running a vision-language model on your own hardware to write a text description for each image, then embedding those descriptions in Postgres with pgvector so you can query by what the picture shows rather than what you happened to name it. The default setup keeps the whole pipeline local, from image-to-text extraction through embedding and retrieval, though you can hand description generation to an OpenAI-compatible vision API while embeddings and search stay on your machine. You can query the index from a web UI, a token-authenticated read-only search API, or a small CLI, and Docker Compose brings the stack up with one command. It is a clean, self-contained example of a semantic search application that teams can adapt to any image corpus, not just memes.