Observed arrival · 2026-09-15
VideoIndex Gives Long Videos a Searchable Memory
An open platform that indexes speech, on-screen text, frames, and scenes so agents can search videos and cite the exact moment they used.
Field notes
The project creates a persistent directory containing synchronized speech, OCR, shot, and frame-embedding data, then exposes that index through agent tools rather than requiring a model to reread the entire video. The page reports 22× real-time indexing on one A100 and distinguishes indexing cost from per-question cost. Its benchmark table covers stratified samples from LVBench and MINERVA, while the roadmap points toward continuously indexed live streams.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue