how would this be different from vector embeddings / semantic search?

Vector embeddings are fuzzy on finding boundaries. With my spec approach, my goal is to get precise start/end times for ffmpeg to do edits. The downside is, that there is a lot of pre-processing of raw footage in my approach. Vectors win on zero-shot flexibility here.