/text-to-image-search

A multimodal search application

Primary LanguagePythonMIT LicenseMIT

Text to Image Search

This service allows to search relevant images with a text prompt. To this end, it uses a pretrained image-and-text encoder model, such as CLIP to encode images and text queries into an n-dimensional vector space. The service then returns the images whose embeddings are closest to the text query's embedding, as measured with the cosine similarity.

Usage

Docker

  • Build the image: 
    docker compose build
  • Start the service(s): 
    docker compose up

Python

  • Install project requirements using 
    pip install -r requirements.txt
  • Prepare image data
  • Start vector database, e.g., with Docker: 
    docker run -p 6333:6333 --mount type=bind,source=$HOME/.qdrant_data,target=/qdrant_data --name vector-db qdrant/qdrant:v1.8.4
  • Start encoding task queue 
    bash deploy/entrypoint.sh task-queue
  • Start HTTP server 
    bash deploy/entrypoint.sh server