Gemini Embedding 2 now allows raw video to be projected directly into a 768-dimensional vector space along with text. This eliminates the need for intermediate text such as transcriptions or frame-by-frame captioning.
Sub-second Video Search Tool Released Using Gemini Native Video Embeddings
This article is a translation. Read the Japanese original
A developer has used this feature to build a CLI tool that indexes hours of footage into ChromaDB, enabling natural language searches and automatic clipping of the relevant sections. Details are available in the demo video on the GitHub README. According to HN Search (backfill), the search speed is reported to be sub-second.
The cost of indexing is approximately $2.50 per hour of video. It is reported that for security camera or sentinel mode footage, operations can be run at a lower cost because idle chunks can be skipped via still-image detection.
Source: Show HN: Gemini can now natively embed video, so I built sub-second video search (HN 438pt, 108 comments) (HN Search (backfill), 2026-03-24)