Computer vision

Video Search

Semantic search across video scenes using natural language.

How it works
  1. Scenes
  2. CLIP embed
  3. Qdrant
  4. Search

The problem

Finding a moment in hours of video means scrubbing. Filenames and tags don’t describe what’s on screen.

How it works

The pipeline detects scene boundaries, embeds each scene with CLIP and indexes the vectors in Qdrant. A FastAPI service and web interface take a natural-language query, embed it the same way and return the matching scenes with timestamps.

03 /

Results

Search latency
200–500 ms