CLIP (Contrastive Language–Image Pre-training) is a technique which efficiently learns visual concepts from natural language supervision. CLIP has found applications in stable diffusion.

This repository aims act as a POC in exploring the ability to use CLIP for video search using natural language outlined in the article found here.

Adapted for a natual language video search engine found here

• libzmq
• python >= 3.8
• go >= 1.18

1. start up the inference zmq server found in the ./inference directory python3 zmq_server.py.
2. start up the go server with go run main.go.

Before running this example, please ensure that your environment is correctly configured and the application is running without errors.

1. index the video clips provided by the examples/videos.

curl -X POST -d '{"videoURI": "<path_to_dir>/examples/videos/<video_name>.mp4" }' https://localhost:3000/insert 

note: it can take a moment for the video to become searchable.

2. then search for a video

curl -X POST -d '{"input": "a man cutting pepper", "maxResults": 1 }' https://localhost:3000/search

• CLI to remove manual setup process
• ability to add dedicated inference machines (currently limited to same host)