A March 24, 2026 Show HN post drew attention because it demonstrated one of the more concrete uses of multimodal embeddings: searching raw video without first turning the footage into text. The project, SentrySearch, packages that idea into a local CLI aimed at dashcam and security video.

According to the repository and the HN write-up, the tool splits footage into overlapping chunks, embeds each chunk directly as video with Gemini Embedding 2, stores the resulting vectors in a local ChromaDB index, and matches a natural-language query against that same embedding space. The top result can then be auto-trimmed back into a clip.

No transcription or frame-captioning stage is required before search.
Default indexing uses 30-second chunks with overlap, plus still-frame skipping so long idle segments do not always get embedded.
The project estimates about $2.50 per hour of footage at default settings, with lower real costs when parked or low-motion footage can be skipped.

That matters because video search projects often fail at the operational layer. They either require a hosted product, a heavyweight vision stack, or manual labeling. SentrySearch instead treats Gemini's video embedding endpoint as an infrastructure primitive and wraps it in something a developer can run locally with Python, ffmpeg, and an API key.

The limitations are also spelled out. Chunk boundaries can miss events that span multiple segments, still-frame detection is heuristic, and Gemini Embedding 2 is still in preview, so both behavior and pricing can move. Even with those caveats, the HN post landed because it translated a new model capability into a workflow that feels immediately usable rather than purely demo-oriented.

Primary source: SentrySearch repository. Community source: Hacker News discussion.

#video-search

Hacker News highlights a practical video-search CLI built on Gemini Embedding 2